Text Generation
GGUF
Portuguese
llama.cpp
quantized
llama-cpp
imatrix
llama
portuguese
european-portuguese
kld-measured
conversational
Instructions to use KikoCis/AMALIA-9B-0626-SFT-GGUF with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- llama.cpp
How to use KikoCis/AMALIA-9B-0626-SFT-GGUF with llama.cpp:
Install (macOS, Linux)
curl -LsSf https://llama.app/install.sh | sh # Start a local OpenAI-compatible server with a web UI: llama serve -hf KikoCis/AMALIA-9B-0626-SFT-GGUF:Q4_K_M # Run inference directly in the terminal: llama cli -hf KikoCis/AMALIA-9B-0626-SFT-GGUF:Q4_K_M
Install from WinGet (Windows)
winget install llama.cpp # Start a local OpenAI-compatible server with a web UI: llama serve -hf KikoCis/AMALIA-9B-0626-SFT-GGUF:Q4_K_M # Run inference directly in the terminal: llama cli -hf KikoCis/AMALIA-9B-0626-SFT-GGUF:Q4_K_M
Use pre-built binary
# Download pre-built binary from: # https://github.com/ggerganov/llama.cpp/releases # Start a local OpenAI-compatible server with a web UI: ./llama-server -hf KikoCis/AMALIA-9B-0626-SFT-GGUF:Q4_K_M # Run inference directly in the terminal: ./llama-cli -hf KikoCis/AMALIA-9B-0626-SFT-GGUF:Q4_K_M
Build from source code
git clone https://github.com/ggerganov/llama.cpp.git cd llama.cpp cmake -B build cmake --build build -j --target llama-server llama-cli # Start a local OpenAI-compatible server with a web UI: ./build/bin/llama-server -hf KikoCis/AMALIA-9B-0626-SFT-GGUF:Q4_K_M # Run inference directly in the terminal: ./build/bin/llama-cli -hf KikoCis/AMALIA-9B-0626-SFT-GGUF:Q4_K_M
Use Docker
docker model run hf.co/KikoCis/AMALIA-9B-0626-SFT-GGUF:Q4_K_M
- LM Studio
- Jan
- vLLM
How to use KikoCis/AMALIA-9B-0626-SFT-GGUF with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "KikoCis/AMALIA-9B-0626-SFT-GGUF" # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:8000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "KikoCis/AMALIA-9B-0626-SFT-GGUF", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }'Use Docker
docker model run hf.co/KikoCis/AMALIA-9B-0626-SFT-GGUF:Q4_K_M
- Ollama
How to use KikoCis/AMALIA-9B-0626-SFT-GGUF with Ollama:
ollama run hf.co/KikoCis/AMALIA-9B-0626-SFT-GGUF:Q4_K_M
- Unsloth Studio
How to use KikoCis/AMALIA-9B-0626-SFT-GGUF with Unsloth Studio:
Install Unsloth Studio (macOS, Linux, WSL)
curl -fsSL https://unsloth.ai/install.sh | sh # Run unsloth studio unsloth studio -H 0.0.0.0 -p 8888 # Then open http://localhost:8888 in your browser # Search for KikoCis/AMALIA-9B-0626-SFT-GGUF to start chatting
Install Unsloth Studio (Windows)
irm https://unsloth.ai/install.ps1 | iex # Run unsloth studio unsloth studio -H 0.0.0.0 -p 8888 # Then open http://localhost:8888 in your browser # Search for KikoCis/AMALIA-9B-0626-SFT-GGUF to start chatting
Using HuggingFace Spaces for Unsloth
# No setup required # Open https://huggingface.co/spaces/unsloth/studio in your browser # Search for KikoCis/AMALIA-9B-0626-SFT-GGUF to start chatting
- Docker Model Runner
How to use KikoCis/AMALIA-9B-0626-SFT-GGUF with Docker Model Runner:
docker model run hf.co/KikoCis/AMALIA-9B-0626-SFT-GGUF:Q4_K_M
- Lemonade
How to use KikoCis/AMALIA-9B-0626-SFT-GGUF with Lemonade:
Pull the model
# Download Lemonade from https://lemonade-server.ai/ lemonade pull KikoCis/AMALIA-9B-0626-SFT-GGUF:Q4_K_M
Run and chat with the model
lemonade run user.AMALIA-9B-0626-SFT-GGUF-Q4_K_M
List all available models
lemonade list
- Atomic Chat
card + metrics + charts + imatrix + Modelfiles
Browse files- .gitattributes +2 -0
- CONTEXT.md +14 -0
- Modelfile.32k +4 -0
- Modelfile.8k +4 -0
- README.md +148 -0
- amalia-9b-pt.imatrix +3 -0
- assemble_card.py +65 -0
- banner.png +3 -0
- metrics/chart-kld-vs-size.png +0 -0
- metrics/chart-top1.png +0 -0
- metrics/quant-summary-ctx512.csv +8 -0
- metrics/quant-summary.csv +9 -0
- reports/artifact-sha256sums.txt +5 -0
- reports/base2048.txt +10 -0
- reports/kld2048-IQ4_XS.txt +91 -0
- reports/kld2048-Q3_K_M.txt +91 -0
- reports/kld2048-Q4_K_M.txt +91 -0
- reports/kld2048-Q5_K_M.txt +91 -0
- reports/kld2048-Q6_K.txt +91 -0
- reports/kld2048-Q8_0.txt +91 -0
- reports/kld2048b-IQ2_M.txt +91 -0
- reports/kld2048b-IQ4_XS.txt +91 -0
- reports/kld2048b-Q3_K_M.txt +91 -0
- reports/kld2048b-Q4_K_M.txt +91 -0
- reports/kld2048b-Q5_K_M.txt +91 -0
- reports/kld2048b-Q6_K.txt +91 -0
- reports/kld2048b-Q8_0.txt +91 -0
- scripts/run_amalia_pt.sh +88 -0
- scripts/run_kld2048.sh +52 -0
.gitattributes
CHANGED
|
@@ -33,3 +33,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
|
| 33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
| 34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
| 35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
|
|
|
| 33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
| 34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
| 35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
| 36 |
+
amalia-9b-pt.imatrix filter=lfs diff=lfs merge=lfs -text
|
| 37 |
+
banner.png filter=lfs diff=lfs merge=lfs -text
|
CONTEXT.md
ADDED
|
@@ -0,0 +1,14 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Context size
|
| 2 |
+
|
| 3 |
+
AMALIA-9B is trained with a native context window of **32768 tokens** (`rope_theta 1e6`).
|
| 4 |
+
|
| 5 |
+
Included Modelfiles set `num_ctx`:
|
| 6 |
+
- `Modelfile.8k` — 8192 (lightest RAM; good for chat)
|
| 7 |
+
- `Modelfile.32k` — 32768 (native maximum)
|
| 8 |
+
|
| 9 |
+
```bash
|
| 10 |
+
ollama create amalia-9b-32k -f Modelfile.32k
|
| 11 |
+
ollama run amalia-9b-32k
|
| 12 |
+
```
|
| 13 |
+
|
| 14 |
+
Larger contexts than 32K require RoPE scaling and will degrade quality — 32K is the honest native limit. KV-cache grows with context, so at 32K budget extra RAM on top of the weights.
|
Modelfile.32k
ADDED
|
@@ -0,0 +1,4 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
FROM ./amalia-9b-Q4_K_M.gguf
|
| 2 |
+
PARAMETER num_ctx 32768
|
| 3 |
+
PARAMETER temperature 0.7
|
| 4 |
+
PARAMETER top_p 0.9
|
Modelfile.8k
ADDED
|
@@ -0,0 +1,4 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
FROM ./amalia-9b-Q4_K_M.gguf
|
| 2 |
+
PARAMETER num_ctx 8192
|
| 3 |
+
PARAMETER temperature 0.7
|
| 4 |
+
PARAMETER top_p 0.9
|
README.md
ADDED
|
@@ -0,0 +1,148 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
base_model:
|
| 4 |
+
- amalia-llm/AMALIA-9B-0626-SFT
|
| 5 |
+
base_model_relation: quantized
|
| 6 |
+
library_name: llama.cpp
|
| 7 |
+
pipeline_tag: text-generation
|
| 8 |
+
tags:
|
| 9 |
+
- gguf
|
| 10 |
+
- quantized
|
| 11 |
+
- llama-cpp
|
| 12 |
+
- imatrix
|
| 13 |
+
- llama
|
| 14 |
+
- portuguese
|
| 15 |
+
- european-portuguese
|
| 16 |
+
- kld-measured
|
| 17 |
+
- conversational
|
| 18 |
+
- endpoints_compatible
|
| 19 |
+
language:
|
| 20 |
+
- pt
|
| 21 |
+
---
|
| 22 |
+
|
| 23 |
+

|
| 24 |
+
|
| 25 |
+
<div style="border:2px solid currentColor; font-family:ui-monospace,'SF Mono','Cascadia Mono',Consolas,monospace;">
|
| 26 |
+
<div style="border-bottom:1px solid currentColor; padding:6px 12px; font-size:11px; letter-spacing:3px; text-transform:uppercase; opacity:0.7; text-align:center;">KIKOCIS // EUROPEAN-PORTUGUESE LLM // IMATRIX GGUF + KLD</div>
|
| 27 |
+
<div style="padding:14px; display:flex; flex-wrap:wrap; align-items:center; justify-content:center; gap:18px;">
|
| 28 |
+
<pre style="margin:0; font-size:9px; line-height:1.15;">
|
| 29 |
+
╔═══════════════╗
|
| 30 |
+
║ A M A L I A ║ pt-PT
|
| 31 |
+
╚═══════════════╝
|
| 32 |
+
┌──┐ ┌──┐ ┌──┐ ┌──┐
|
| 33 |
+
│Q3│ │Q4│ │Q5│ │Q8│ ← ladder
|
| 34 |
+
└──┘ └──┘ └──┘ └──┘
|
| 35 |
+
●────●────●────●
|
| 36 |
+
KLD vs F16
|
| 37 |
+
</pre>
|
| 38 |
+
<div style="text-align:center;">
|
| 39 |
+
<div style="font-size:23px; font-weight:800; letter-spacing:1px;">AMALIA-9B · GGUF</div>
|
| 40 |
+
<div style="font-size:12.5px; letter-spacing:1px; opacity:0.8; margin-top:5px;"><span style="white-space:nowrap;">llama · 9B</span> · <span style="white-space:nowrap;">32K ctx</span> · <span style="white-space:nowrap;">imatrix (pt)</span> · <span style="white-space:nowrap;">KLD-measured</span></div>
|
| 41 |
+
</div>
|
| 42 |
+
</div>
|
| 43 |
+
<table style="display:table; table-layout:fixed; width:100%; margin:0; border-collapse:collapse; font-size:12px;">
|
| 44 |
+
<tr>
|
| 45 |
+
<td style="border-top:1px solid currentColor; border-right:1px solid currentColor; padding:8px 12px;"><div style="font-size:10px; letter-spacing:1px; opacity:0.6;">FORMAT</div><div style="font-weight:700;">GGUF (imatrix)</div></td>
|
| 46 |
+
<td style="border-top:1px solid currentColor; border-right:1px solid currentColor; padding:8px 12px;"><div style="font-size:10px; letter-spacing:1px; opacity:0.6;">SIZES</div><div style="font-weight:700;">3.1 – 9.1 GB</div></td>
|
| 47 |
+
<td style="border-top:1px solid currentColor; border-right:1px solid currentColor; padding:8px 12px;"><div style="font-size:10px; letter-spacing:1px; opacity:0.6;">ARCH</div><div style="font-weight:700;">Llama · 42L</div></td>
|
| 48 |
+
<td style="border-top:1px solid currentColor; padding:8px 12px;"><div style="font-size:10px; letter-spacing:1px; opacity:0.6;">CONTEXT</div><div style="font-weight:700;">32768</div></td>
|
| 49 |
+
</tr>
|
| 50 |
+
<tr>
|
| 51 |
+
<td style="border-top:1px solid currentColor; border-right:1px solid currentColor; padding:8px 12px;"><div style="font-size:10px; letter-spacing:1px; opacity:0.6;">IMATRIX</div><div style="font-weight:700;">pt-PT corpus</div></td>
|
| 52 |
+
<td style="border-top:1px solid currentColor; border-right:1px solid currentColor; padding:8px 12px;"><div style="font-size:10px; letter-spacing:1px; opacity:0.6;">VALIDATION</div><div style="font-weight:700;">KLD vs F16</div></td>
|
| 53 |
+
<td style="border-top:1px solid currentColor; border-right:1px solid currentColor; padding:8px 12px;"><div style="font-size:10px; letter-spacing:1px; opacity:0.6;">LANGUAGE</div><div style="font-weight:700;">Português (EU)</div></td>
|
| 54 |
+
<td style="border-top:1px solid currentColor; padding:8px 12px;"><div style="font-size:10px; letter-spacing:1px; opacity:0.6;">LICENSE</div><div style="font-weight:700;">Apache-2.0</div></td>
|
| 55 |
+
</tr>
|
| 56 |
+
</table>
|
| 57 |
+
</div>
|
| 58 |
+
|
| 59 |
+
# AMALIA-9B-0626-SFT — GGUF (imatrix + KLD)
|
| 60 |
+
> imatrix-quantized GGUFs of **AMALIA-9B**, an open **European-Portuguese** instruction model — one of the first open LLMs developed in **Portugal**. Runs from **~3.1 GB** (IQ2_M) to **~9.1 GB** (Q8_0). Calibrated (imatrix) **on Portuguese text**, and every quant is measured against the full-precision model with **KLD** so you can pick a size with the fidelity cost in front of you. Credit: this is **the AMALIA-LLM project's** model — [`amalia-llm/AMALIA-9B-0626-SFT`](https://huggingface.co/amalia-llm/AMALIA-9B-0626-SFT); ours is the quant ladder + metrics.
|
| 61 |
+
|
| 62 |
+
## ✅ Recommended files
|
| 63 |
+
| Use case | File | Notes |
|
| 64 |
+
|---|---|---|
|
| 65 |
+
| ⭐ Best value | `amalia-9b-Q5_K_M.gguf` | The pick — **lower KLD than Q6_K** at 1 GB less (0.051 vs 0.078). ~6.1 GB. |
|
| 66 |
+
| Safe default | `amalia-9b-Q4_K_M.gguf` | Standard K-quant, ~5.2 GB; noticeably less faithful than Q5 (KLD 0.22 vs 0.05). |
|
| 67 |
+
| Max fidelity | `amalia-9b-Q8_0.gguf` | Lowest KLD (0.006) — near-lossless. ~9.1 GB. |
|
| 68 |
+
| Tightest RAM | `amalia-9b-Q3_K_M.gguf` | ~4.2 GB — runs on 8 GB machines, still coherent. |
|
| 69 |
+
| Smallest (aggressive) | `amalia-9b-IQ2_M.gguf` | ~3.1 GB — fits very tight RAM, at a real quality cost (top-1 61%). |
|
| 70 |
+
|
| 71 |
+
## 📦 Files (the full ladder)
|
| 72 |
+
| Quant | Bits | File size | Notes |
|
| 73 |
+
|---|---|---:|---|
|
| 74 |
+
| IQ2_M | ~2.7 | ~3.1 GB | Smallest — aggressive i-quant, real quality trade-off. |
|
| 75 |
+
| Q3_K_M | 3 | ~4.2 GB | Smallest *comfortable* quant. |
|
| 76 |
+
| IQ4_XS | ~4.3 | ~4.7 GB | Compact i-quant (imatrix). |
|
| 77 |
+
| Q4_K_M | 4 | ~5.2 GB | Standard K-quant. |
|
| 78 |
+
| Q5_K_M | 5 | ~6.1 GB | ⭐ Best quality/size. |
|
| 79 |
+
| Q6_K | 6 | ~7.0 GB | Near-lossless. |
|
| 80 |
+
| Q8_0 | 8 | ~9.1 GB | Highest precision. |
|
| 81 |
+
<!-- K-quants (Q*_K) = solid general quants. I-quants (IQ*) = smaller, need an imatrix (shipped). Higher number = more bits = closer to the original + bigger. -->
|
| 82 |
+
|
| 83 |
+
## 📊 Metrics — objective quality vs the F16 reference
|
| 84 |
+
**KLD** (Kullback–Leibler divergence, nats) measures how far each quant's output distribution drifts from the full-precision model — the gold-standard fidelity metric (lower = closer). **Top-1 match** = how often the quant's top token agrees with F16. Measured with `llama-perplexity --kl-divergence` over a **Portuguese** corpus at **ctx 2048**, vs the F16 GGUF.
|
| 85 |
+
|
| 86 |
+
| Model | Size GB | PPL | ΔPPL vs F16 | KLD mean | KLD p95 | Top-1 match |
|
| 87 |
+
|---|---:|---:|---:|---:|---:|---:|
|
| 88 |
+
| **F16 (reference)** | 17.1 | 20.04 | 0.000 | 0.0000 | 0.0000 | 100.0% |
|
| 89 |
+
| Q8_0 | 9.06 | 20.20 | +0.177 | 0.0057 | 0.0198 | 96.38% |
|
| 90 |
+
| Q6_K | 7.00 | 19.93 | -0.091 | 0.0779 | 0.1740 | 89.06% |
|
| 91 |
+
| Q5_K_M | 6.08 | 19.80 | -0.227 | 0.0511 | 0.1747 | 88.89% |
|
| 92 |
+
| Q4_K_M | 5.20 | 20.23 | +0.203 | 0.2230 | 0.7429 | 81.39% |
|
| 93 |
+
| IQ4_XS | 4.70 | 20.78 | +0.757 | 0.2173 | 0.7600 | 80.55% |
|
| 94 |
+
| Q3_K_M | 4.25 | 22.24 | +2.215 | 0.3258 | 1.1486 | 73.58% |
|
| 95 |
+
| IQ2_M | 3.11 | 21.99 | +1.968 | 0.8211 | 2.9052 | 60.69% |
|
| 96 |
+
|
| 97 |
+
<sub>Full per-quant reports in [`reports/`](reports); machine-readable summary in [`metrics/quant-summary.csv`](metrics/quant-summary.csv); SHA-256 of every file in [`reports/artifact-sha256sums.txt`](reports/artifact-sha256sums.txt). Note: this is an **instruction-tuned** model measured on **raw** text, so the absolute PPL is higher than a base model's — what matters here is the **relative** KLD/top-1 across quants.</sub>
|
| 98 |
+
|
| 99 |
+
> **Why does Q5_K_M beat Q6_K on KLD here?** It's real and reproducible for this model: Q5_K_M's imatrix-weighted bit allocation is very effective, so it lands closer to F16 on mean-KLD than Q6_K (which stays marginally ahead on top-1). Both are near-lossless — so **Q5_K_M is the better value**, and Q8_0 is the true fidelity ceiling.
|
| 100 |
+
|
| 101 |
+
> **A note on i-quants:** we tested the i-quants too. For *this* model the K-quants win at low bit — e.g. IQ3_M came out *worse* than Q3_K_M on KLD, so it isn't shipped. **IQ2_M is included** as the smallest option for very tight RAM, but with an honest quality cost (top-1 61%); if you can spare the space, Q3_K_M is a big step up.
|
| 102 |
+
|
| 103 |
+
### 📈 Charts
|
| 104 |
+

|
| 105 |
+

|
| 106 |
+
|
| 107 |
+
## 🧮 Will it fit? (RAM/VRAM cheat-sheet)
|
| 108 |
+
Memory ≈ weights + KV-cache (KV grows with context). Rough guide:
|
| 109 |
+
| you have | comfortable quant | context |
|
| 110 |
+
|---|---|---|
|
| 111 |
+
| 6 GB | IQ2_M / Q3_K_M | ~8–16K |
|
| 112 |
+
| 8 GB | Q3_K_M / Q4_K_M | ~16–32K |
|
| 113 |
+
| 12 GB | Q5_K_M / Q6_K | ~32K |
|
| 114 |
+
| 16 GB+ | Q6_K / Q8_0 | ~32K (native) |
|
| 115 |
+
|
| 116 |
+
## 🚀 How to run it
|
| 117 |
+
```bash
|
| 118 |
+
# ollama (pulls straight from this repo)
|
| 119 |
+
ollama run hf.co/KikoCis/AMALIA-9B-0626-SFT-GGUF:Q5_K_M
|
| 120 |
+
|
| 121 |
+
# llama.cpp
|
| 122 |
+
llama-server -m amalia-9b-Q5_K_M.gguf -c 32768 --jinja
|
| 123 |
+
llama-cli -m amalia-9b-Q5_K_M.gguf -p "Explica em português o que é a saudade." -ngl 99
|
| 124 |
+
```
|
| 125 |
+
**Recommended sampling**: temperature ~0.7, top_p ~0.9. It's a **chat/instruct** model (uses its built-in chat template) — prompt it in **European Portuguese** for best results.
|
| 126 |
+
|
| 127 |
+
## ⚠️ Good to know
|
| 128 |
+
- **Strengths**: European-Portuguese fluency and instruction-following; one of the first open LLMs built in Portugal.
|
| 129 |
+
- **Limits**: specialised for Portuguese — not a coding/agentic model, and not tuned for English. Low-bit quants (IQ2/Q3) trade real fidelity for size (see KLD).
|
| 130 |
+
- Absolute perplexity looks high only because it's measured on raw text with an instruct model; the KLD table is the fair quant-quality signal.
|
| 131 |
+
|
| 132 |
+
## 📊 Evaluation methodology (how the numbers were measured)
|
| 133 |
+
- **What**: quantization **fidelity** vs the F16 GGUF — not a task benchmark. `llama-perplexity --kl-divergence` (KLD mean/p95/max, ΔPPL, top-1 agreement).
|
| 134 |
+
- **Corpus**: European-Portuguese text (Wikipedia + diverse), **ctx 2048**, same corpus family used for the imatrix calibration.
|
| 135 |
+
- **Reference**: the model's own **F16** GGUF (KLD = 0 by definition).
|
| 136 |
+
- **Date**: 2026-07. *Caveat: relative fidelity ranking across quants of this model; not comparable across different models/corpora.*
|
| 137 |
+
|
| 138 |
+
## 🔁 Provenance & reproducibility
|
| 139 |
+
- **Scripts**: [`scripts/`](scripts) — exact convert → imatrix → quant → KLD commands.
|
| 140 |
+
- **imatrix**: [`amalia-9b-pt.imatrix`](amalia-9b-pt.imatrix) — importance matrix (Portuguese calibration) used for the IQ/K quants.
|
| 141 |
+
- **Checksums**: [`reports/artifact-sha256sums.txt`](reports/artifact-sha256sums.txt) — `shasum -a 256 -c` to verify.
|
| 142 |
+
- **Source**: [`amalia-llm/AMALIA-9B-0626-SFT`](https://huggingface.co/amalia-llm/AMALIA-9B-0626-SFT) — weights **unmodified** (faithful quantization).
|
| 143 |
+
|
| 144 |
+
## 🗒️ Changelog
|
| 145 |
+
- 2026-07 v1: imatrix GGUF ladder (IQ2_M → Q8_0) + KLD/PPL metrics, Portuguese-calibrated.
|
| 146 |
+
|
| 147 |
+
## 📚 Credit & license
|
| 148 |
+
Model, weights, training data: **© the AMALIA-LLM project** — [model](https://huggingface.co/amalia-llm/AMALIA-9B-0626-SFT) · [website](https://amaliallm.pt/) · [paper (PROPOR 2026)](https://aclanthology.org/2026.propor-1.38/). Quant ladder + imatrix (pt) + KLD/PPL metrics: KikoCis. **Apache-2.0** (same as upstream). No weights modified.
|
amalia-9b-pt.imatrix
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:1a919543f915131aca29bfd62b609ffc664a6e4d580921630558f6c21da87394
|
| 3 |
+
size 6238400
|
assemble_card.py
ADDED
|
@@ -0,0 +1,65 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/usr/bin/env python3
|
| 2 |
+
"""Fill the KLD table into README.md and render the two charts, from quant-summary.csv."""
|
| 3 |
+
import csv, os, sys
|
| 4 |
+
import matplotlib
|
| 5 |
+
matplotlib.use("Agg")
|
| 6 |
+
import matplotlib.pyplot as plt
|
| 7 |
+
|
| 8 |
+
W = "/Users/kikocisneros/coco_ppl/aq_amalia"
|
| 9 |
+
PUB = f"{W}/publish"
|
| 10 |
+
CSV = f"{W}/metrics/quant-summary.csv"
|
| 11 |
+
os.makedirs(f"{PUB}/metrics", exist_ok=True)
|
| 12 |
+
|
| 13 |
+
rows = list(csv.DictReader(open(CSV)))
|
| 14 |
+
def f(x, d=None):
|
| 15 |
+
try: return float(x)
|
| 16 |
+
except: return d
|
| 17 |
+
|
| 18 |
+
# order big->small by size for the table
|
| 19 |
+
order = sorted(rows, key=lambda r: -f(r["size_gb"], 0))
|
| 20 |
+
|
| 21 |
+
# --- markdown KLD table ---
|
| 22 |
+
hdr = "| Model | Size GB | PPL | ΔPPL vs F16 | KLD mean | KLD p95 | Top-1 match |\n|---|---:|---:|---:|---:|---:|---:|"
|
| 23 |
+
lines = [hdr]
|
| 24 |
+
for r in order:
|
| 25 |
+
name = r["name"]
|
| 26 |
+
ppl = r["ppl"]; dppl = r["ppl_delta_vs_f16"]
|
| 27 |
+
km = r["kld_mean"]; kp = r["kld_p95"]; t1 = r["top1_match_pct"]
|
| 28 |
+
if name == "F16":
|
| 29 |
+
lines.append(f"| **F16 (reference)** | {f(r['size_gb']):.1f} | {f(ppl):.2f} | 0.000 | 0.0000 | 0.0000 | 100.0% |")
|
| 30 |
+
else:
|
| 31 |
+
dv = f(dppl, 0); sign = "+" if dv >= 0 else ""
|
| 32 |
+
lines.append(f"| {name} | {f(r['size_gb']):.2f} | {f(ppl):.2f} | {sign}{dv:.3f} | {f(km):.4f} | {f(kp):.4f} | {f(t1):.2f}% |")
|
| 33 |
+
table = "\n".join(lines)
|
| 34 |
+
|
| 35 |
+
readme = open(f"{PUB}/README.md").read()
|
| 36 |
+
readme = readme.replace("<!--KLD_TABLE_PLACEHOLDER-->", table)
|
| 37 |
+
open(f"{PUB}/README.md", "w").write(readme)
|
| 38 |
+
print("README table filled:\n" + table)
|
| 39 |
+
|
| 40 |
+
# --- chart 1: KLD mean vs size ---
|
| 41 |
+
q = [r for r in order if r["name"] != "F16"]
|
| 42 |
+
q = sorted(q, key=lambda r: f(r["size_gb"]))
|
| 43 |
+
sizes = [f(r["size_gb"]) for r in q]; klds = [f(r["kld_mean"]) for r in q]; names = [r["name"] for r in q]
|
| 44 |
+
plt.figure(figsize=(7,4.2))
|
| 45 |
+
plt.plot(sizes, klds, "o-", color="#2ca02c", lw=2, ms=7)
|
| 46 |
+
for x,y,n in zip(sizes,klds,names):
|
| 47 |
+
plt.annotate(n, (x,y), textcoords="offset points", xytext=(6,6), fontsize=8)
|
| 48 |
+
plt.yscale("log"); plt.xlabel("file size (GB)"); plt.ylabel("KLD mean (nats, log) — lower = closer to F16")
|
| 49 |
+
plt.title("AMALIA-9B GGUF — fidelity vs size (KLD vs F16, pt, ctx 2048)")
|
| 50 |
+
plt.grid(True, which="both", alpha=0.3); plt.tight_layout()
|
| 51 |
+
plt.savefig(f"{PUB}/metrics/chart-kld-vs-size.png", dpi=130); plt.close()
|
| 52 |
+
|
| 53 |
+
# --- chart 2: top-1 match ---
|
| 54 |
+
plt.figure(figsize=(7,4.2))
|
| 55 |
+
t1s = [f(r["top1_match_pct"]) for r in q]
|
| 56 |
+
bars = plt.bar(names, t1s, color="#1f77b4")
|
| 57 |
+
for b,v in zip(bars,t1s): plt.text(b.get_x()+b.get_width()/2, v+0.4, f"{v:.1f}", ha="center", fontsize=8)
|
| 58 |
+
plt.ylabel("Top-1 token match vs F16 (%)"); plt.ylim(min(t1s)-3, 100.5)
|
| 59 |
+
plt.title("AMALIA-9B GGUF — top-1 agreement with F16")
|
| 60 |
+
plt.grid(True, axis="y", alpha=0.3); plt.tight_layout()
|
| 61 |
+
plt.savefig(f"{PUB}/metrics/chart-top1.png", dpi=130); plt.close()
|
| 62 |
+
|
| 63 |
+
# copy the csv into publish
|
| 64 |
+
import shutil; shutil.copy(CSV, f"{PUB}/metrics/quant-summary.csv")
|
| 65 |
+
print("charts + csv written to", f"{PUB}/metrics/")
|
banner.png
ADDED
|
Git LFS Details
|
metrics/chart-kld-vs-size.png
ADDED
|
metrics/chart-top1.png
ADDED
|
metrics/quant-summary-ctx512.csv
ADDED
|
@@ -0,0 +1,8 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
name,size_gb,ppl,ppl_delta_vs_f16,kld_mean,kld_p95,kld_max,top1_match_pct
|
| 2 |
+
F16,17.05,4.44,0.0,0.0,0.0,0.0,100.0
|
| 3 |
+
Q8_0,9.06,32.115409,0.091330,0.009722,0.031179,3.845082,95.497
|
| 4 |
+
Q6_K,7.00,31.491478,-0.532602,0.088915,0.260961,16.998377,86.758
|
| 5 |
+
Q5_K_M,6.07,31.717572,-0.306507,0.080225,0.280754,6.408488,86.025
|
| 6 |
+
Q4_K_M,5.20,32.896111,0.872032,0.206477,0.734504,13.929957,78.730
|
| 7 |
+
IQ4_XS,4.70,30.450651,-1.573428,0.384149,1.645543,30.692392,76.901
|
| 8 |
+
Q3_K_M,4.24,35.781539,3.757460,0.420631,1.423969,32.735237,70.642
|
metrics/quant-summary.csv
ADDED
|
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
name,size_gb,ppl,ppl_delta_vs_f16,kld_mean,kld_p95,kld_max,top1_match_pct
|
| 2 |
+
F16,17.05,20.0376,0.0,0.0,0.0,0.0,100.0
|
| 3 |
+
Q8_0,9.06,20.200587,0.177055,0.005708,0.019779,1.475488,96.381
|
| 4 |
+
Q6_K,7.00,19.932671,-0.090860,0.077855,0.174014,16.764179,89.056
|
| 5 |
+
Q5_K_M,6.08,19.797018,-0.226513,0.051125,0.174651,9.869919,88.887
|
| 6 |
+
Q4_K_M,5.20,20.226883,0.203352,0.222994,0.742914,21.454760,81.387
|
| 7 |
+
IQ4_XS,4.70,20.780607,0.757076,0.217275,0.760046,17.964737,80.554
|
| 8 |
+
Q3_K_M,4.25,22.238185,2.214654,0.325828,1.148634,19.564314,73.580
|
| 9 |
+
IQ2_M,3.11,21.991922,1.968390,0.821109,2.905154,27.317856,60.690
|
reports/artifact-sha256sums.txt
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
c653273b579c7a8ae813389db68178f3520fff2735aee6a4799f194ddef29b6d amalia-9b-IQ2_M.gguf
|
| 2 |
+
6111748cecc0d26478194412cc34f0eb63f91b1b62098755298019f868b503b1 amalia-9b-IQ4_XS.gguf
|
| 3 |
+
2187c384a1d879825e3fab47867530093765dd5b1509393b179e73dc332302ab amalia-9b-Q3_K_M.gguf
|
| 4 |
+
e1be526dea752b1cd8e019d0b6444c9cb76f812a37e1f9bc309360923d6a1718 amalia-9b-Q4_K_M.gguf
|
| 5 |
+
9487ae725635c5ca6b7c4e3768ca8bd9e28adcb2e88ff0a54d6e87250e999368 amalia-9b-Q5_K_M.gguf
|
reports/base2048.txt
ADDED
|
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
0.07.412.398 I
|
| 2 |
+
0.07.412.417 I system_info: n_threads = 12 (n_threads_batch = 12) / 16 | MTL : EMBED_LIBRARY = 1 | CPU : NEON = 1 | ARM_FMA = 1 | FP16_VA = 1 | MATMUL_INT8 = 1 | DOTPROD = 1 | SME = 1 | ACCELERATE = 1 | OPENMP = 1 | REPACK = 1 |
|
| 3 |
+
0.07.412.501 I perplexity: saving all logits to /Users/kikocisneros/coco_ppl/aq_amalia/kld_base_2048.dat
|
| 4 |
+
0.07.412.504 I perplexity: tokenizing the input ..
|
| 5 |
+
0.07.557.142 I perplexity: tokenization took 144.636 ms
|
| 6 |
+
0.07.557.200 I perplexity: calculating perplexity over 44 chunks, n_ctx=2048, batch_size=2048, n_seq=1
|
| 7 |
+
0.10.211.335 I perplexity: 2.63 seconds per pass - ETA 1.92 minutes
|
| 8 |
+
[1]39.2953,[2]43.3171,[3]30.9450,[4]21.2218,[5]16.7458,[6]18.9088,[7]21.3953,[8]23.6025,[9]20.7184,[10]23.2438,[11]20.7743,[12]19.9841,[13]19.8370,[14]21.0076,[15]20.2485,[16]18.7585,[17]18.6742,[18]19.2116,[19]19.3664,[20]19.7867,[21]19.9371,[22]19.1936,[23]18.1903,[24]18.3383,[25]17.6663,[26]17.9583,[27]17.1198,[28]17.3306,[29]17.6339,[30]18.0336,[31]17.8652,[32]17.6116,[33]17.8595,[34]17.9006,[35]18.1363,[36]18.6202,[37]18.8459,[38]19.2546,[39]19.4834,[40]19.5976,[41]19.8347,[42]20.0916,[43]20.3915,[44]20.0376,
|
| 9 |
+
3.42.832.788 I Final estimate: PPL = 20.0376 +/- 0.28573
|
| 10 |
+
|
reports/kld2048-IQ4_XS.txt
ADDED
|
@@ -0,0 +1,91 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
0.02.265.152 I
|
| 2 |
+
0.02.265.171 I system_info: n_threads = 12 (n_threads_batch = 12) / 16 | MTL : EMBED_LIBRARY = 1 | CPU : NEON = 1 | ARM_FMA = 1 | FP16_VA = 1 | MATMUL_INT8 = 1 | DOTPROD = 1 | SME = 1 | ACCELERATE = 1 | OPENMP = 1 | REPACK = 1 |
|
| 3 |
+
0.02.282.825 I kl_divergence: computing over 44 chunks, n_ctx=2048, batch_size=2048, n_seq=1
|
| 4 |
+
0.06.574.032 I kl_divergence: 4.29 seconds per pass - ETA 3.13 minutes
|
| 5 |
+
|
| 6 |
+
chunk PPL ln(PPL(Q)/PPL(base)) KL Divergence Δp RMS Same top p
|
| 7 |
+
1 40.3335 ± 3.9997 0.02855 ± 0.02025 0.13629 ± 0.00600 8.309 ± 0.525 % 77.126 ± 1.314 %
|
| 8 |
+
2 45.6377 ± 3.2447 0.05342 ± 0.01545 0.14017 ± 0.00452 8.435 ± 0.420 % 77.859 ± 0.918 %
|
| 9 |
+
3 32.5277 ± 1.9038 0.05070 ± 0.01138 0.11585 ± 0.00354 7.950 ± 0.385 % 81.167 ± 0.706 %
|
| 10 |
+
4 22.3582 ± 1.1201 0.05278 ± 0.00981 0.10672 ± 0.00309 7.709 ± 0.320 % 83.431 ± 0.581 %
|
| 11 |
+
5 17.5481 ± 0.7691 0.04729 ± 0.00824 0.09422 ± 0.00274 7.386 ± 0.299 % 85.357 ± 0.494 %
|
| 12 |
+
6 19.9353 ± 0.8053 0.05327 ± 0.00787 0.10348 ± 0.00253 7.645 ± 0.260 % 84.229 ± 0.465 %
|
| 13 |
+
7 22.4908 ± 0.8507 0.05041 ± 0.00738 0.10800 ± 0.00232 7.619 ± 0.231 % 83.215 ± 0.442 %
|
| 14 |
+
8 24.8056 ± 0.8778 0.05013 ± 0.00704 0.11587 ± 0.00227 7.686 ± 0.216 % 82.270 ± 0.422 %
|
| 15 |
+
9 21.6580 ± 0.7181 0.04474 ± 0.00654 0.11110 ± 0.00216 7.668 ± 0.205 % 83.089 ± 0.391 %
|
| 16 |
+
10 24.0176 ± 0.7624 0.03313 ± 0.00628 0.11687 ± 0.00205 7.632 ± 0.191 % 82.356 ± 0.377 %
|
| 17 |
+
11 21.4310 ± 0.6462 0.03173 ± 0.00596 0.11504 ± 0.00200 7.738 ± 0.182 % 83.009 ± 0.354 %
|
| 18 |
+
12 20.6137 ± 0.5894 0.03158 ± 0.00565 0.11387 ± 0.00188 7.819 ± 0.170 % 83.187 ± 0.338 %
|
| 19 |
+
13 20.3602 ± 0.5565 0.02655 ± 0.00542 0.11427 ± 0.00178 7.884 ± 0.161 % 83.021 ± 0.326 %
|
| 20 |
+
14 21.6156 ± 0.5723 0.02909 ± 0.00524 0.11711 ± 0.00173 7.871 ± 0.154 % 82.733 ± 0.316 %
|
| 21 |
+
15 19.3184 ± 0.4889 -0.04649 ± 0.00697 0.24254 ± 0.00679 12.238 ± 0.256 % 80.932 ± 0.317 %
|
| 22 |
+
16 18.0005 ± 0.4389 -0.04076 ± 0.00660 0.23150 ± 0.00639 12.002 ± 0.246 % 81.488 ± 0.304 %
|
| 23 |
+
17 18.1333 ± 0.4253 -0.02893 ± 0.00634 0.22577 ± 0.00603 11.875 ± 0.235 % 81.375 ± 0.295 %
|
| 24 |
+
18 18.7734 ± 0.4274 -0.02264 ± 0.00605 0.21874 ± 0.00570 11.668 ± 0.227 % 81.378 ± 0.287 %
|
| 25 |
+
19 19.0545 ± 0.4212 -0.01582 ± 0.00583 0.21391 ± 0.00541 11.529 ± 0.218 % 81.335 ± 0.279 %
|
| 26 |
+
20 19.5749 ± 0.4208 -0.01036 ± 0.00563 0.21026 ± 0.00515 11.357 ± 0.211 % 81.271 ± 0.273 %
|
| 27 |
+
21 19.7828 ± 0.4143 -0.00736 ± 0.00541 0.20494 ± 0.00491 11.190 ± 0.204 % 81.362 ± 0.266 %
|
| 28 |
+
22 19.0823 ± 0.3883 -0.00542 ± 0.00520 0.19826 ± 0.00469 11.031 ± 0.198 % 81.707 ± 0.258 %
|
| 29 |
+
23 18.1130 ± 0.3588 -0.00388 ± 0.00502 0.19222 ± 0.00450 10.905 ± 0.193 % 82.082 ± 0.250 %
|
| 30 |
+
24 18.3864 ± 0.3579 0.00334 ± 0.00491 0.19097 ± 0.00433 10.892 ± 0.187 % 81.953 ± 0.245 %
|
| 31 |
+
25 17.7117 ± 0.3365 0.00325 ± 0.00474 0.18507 ± 0.00416 10.719 ± 0.183 % 82.311 ± 0.239 %
|
| 32 |
+
26 18.0524 ± 0.3363 0.00589 ± 0.00461 0.18246 ± 0.00401 10.646 ± 0.178 % 82.209 ± 0.234 %
|
| 33 |
+
27 17.2297 ± 0.3139 0.00719 ± 0.00448 0.17912 ± 0.00388 10.595 ± 0.173 % 82.448 ± 0.229 %
|
| 34 |
+
28 17.5078 ± 0.3130 0.01093 ± 0.00437 0.17648 ± 0.00375 10.505 ± 0.169 % 82.488 ± 0.225 %
|
| 35 |
+
29 17.8562 ± 0.3150 0.01326 ± 0.00430 0.17700 ± 0.00363 10.494 ± 0.165 % 82.317 ± 0.222 %
|
| 36 |
+
30 18.2704 ± 0.3167 0.01375 ± 0.00422 0.17611 ± 0.00352 10.429 ± 0.161 % 82.199 ± 0.218 %
|
| 37 |
+
31 18.1377 ± 0.3079 0.01583 ± 0.00412 0.17346 ± 0.00341 10.356 ± 0.157 % 82.304 ± 0.214 %
|
| 38 |
+
32 18.4414 ± 0.3083 0.04679 ± 0.00471 0.21744 ± 0.00391 12.102 ± 0.173 % 81.137 ± 0.216 %
|
| 39 |
+
33 18.7050 ± 0.3074 0.04699 ± 0.00460 0.21434 ± 0.00379 11.998 ± 0.169 % 81.089 ± 0.213 %
|
| 40 |
+
34 18.7511 ± 0.3029 0.04714 ± 0.00449 0.21119 ± 0.00368 11.880 ± 0.166 % 81.045 ± 0.210 %
|
| 41 |
+
35 19.0228 ± 0.3036 0.04844 ± 0.00441 0.20948 �� 0.00359 11.800 ± 0.162 % 81.022 ± 0.207 %
|
| 42 |
+
36 19.5390 ± 0.3089 0.04896 ± 0.00435 0.20948 ± 0.00350 11.746 ± 0.159 % 80.854 ± 0.205 %
|
| 43 |
+
37 19.7770 ± 0.3082 0.04900 ± 0.00425 0.20617 ± 0.00340 11.636 ± 0.157 % 80.954 ± 0.202 %
|
| 44 |
+
38 20.2114 ± 0.3112 0.04925 ± 0.00417 0.20401 ± 0.00332 11.547 ± 0.154 % 80.905 ± 0.199 %
|
| 45 |
+
39 20.4638 ± 0.3109 0.04984 ± 0.00408 0.20165 ± 0.00323 11.452 ± 0.151 % 80.851 ± 0.197 %
|
| 46 |
+
40 20.5837 ± 0.3085 0.04982 ± 0.00401 0.19926 ± 0.00316 11.368 ± 0.149 % 80.880 ± 0.194 %
|
| 47 |
+
41 20.8094 ± 0.3084 0.04872 ± 0.00394 0.19737 ± 0.00308 11.293 ± 0.146 % 80.876 ± 0.192 %
|
| 48 |
+
42 21.0830 ± 0.3090 0.04890 ± 0.00388 0.19638 ± 0.00301 11.238 ± 0.144 % 80.836 ± 0.190 %
|
| 49 |
+
43 21.3715 ± 0.3096 0.04766 ± 0.00381 0.19445 ± 0.00295 11.163 ± 0.142 % 80.820 ± 0.188 %
|
| 50 |
+
44 20.9837 ± 0.2996 0.04684 ± 0.00373 0.19052 ± 0.00288 11.048 ± 0.140 % 81.061 ± 0.185 %
|
| 51 |
+
|
| 52 |
+
====== Perplexity statistics ======
|
| 53 |
+
Mean PPL(Q) : 20.983665 ± 0.299595
|
| 54 |
+
Mean PPL(base) : 20.023532 ± 0.285215
|
| 55 |
+
Cor(ln(PPL(Q)), ln(PPL(base))): 96.57%
|
| 56 |
+
Mean ln(PPL(Q)/PPL(base)) : 0.046836 ± 0.003733
|
| 57 |
+
Mean PPL(Q)/PPL(base) : 1.047950 ± 0.003912
|
| 58 |
+
Mean PPL(Q)-PPL(base) : 0.960133 ± 0.077854
|
| 59 |
+
|
| 60 |
+
====== KL divergence statistics ======
|
| 61 |
+
Mean KLD: 0.190517 ± 0.002882
|
| 62 |
+
Maximum KLD: 15.982407
|
| 63 |
+
99.9% KLD: 8.710998
|
| 64 |
+
99.0% KLD: 2.718909
|
| 65 |
+
95.0% KLD: 0.599236
|
| 66 |
+
90.0% KLD: 0.322753
|
| 67 |
+
Median KLD: 0.065426
|
| 68 |
+
10.0% KLD: 0.003866
|
| 69 |
+
5.0% KLD: 0.000702
|
| 70 |
+
1.0% KLD: 0.000018
|
| 71 |
+
0.1% KLD: 0.000000
|
| 72 |
+
Minimum KLD: -0.000021
|
| 73 |
+
|
| 74 |
+
====== Token probability statistics ======
|
| 75 |
+
Mean Δp: -0.745 ± 0.052 %
|
| 76 |
+
Maximum Δp: 99.976%
|
| 77 |
+
99.9% Δp: 90.469%
|
| 78 |
+
99.0% Δp: 30.855%
|
| 79 |
+
95.0% Δp: 10.810%
|
| 80 |
+
90.0% Δp: 5.319%
|
| 81 |
+
75.0% Δp: 0.652%
|
| 82 |
+
Median Δp: -0.007%
|
| 83 |
+
25.0% Δp: -1.643%
|
| 84 |
+
10.0% Δp: -7.913%
|
| 85 |
+
5.0% Δp: -14.614%
|
| 86 |
+
1.0% Δp: -38.569%
|
| 87 |
+
0.1% Δp: -92.396%
|
| 88 |
+
Minimum Δp: -99.956%
|
| 89 |
+
RMS Δp : 11.048 ± 0.140 %
|
| 90 |
+
Same top p: 81.061 ± 0.185 %
|
| 91 |
+
|
reports/kld2048-Q3_K_M.txt
ADDED
|
@@ -0,0 +1,91 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
0.02.111.565 I
|
| 2 |
+
0.02.111.584 I system_info: n_threads = 12 (n_threads_batch = 12) / 16 | MTL : EMBED_LIBRARY = 1 | CPU : NEON = 1 | ARM_FMA = 1 | FP16_VA = 1 | MATMUL_INT8 = 1 | DOTPROD = 1 | SME = 1 | ACCELERATE = 1 | OPENMP = 1 | REPACK = 1 |
|
| 3 |
+
0.02.126.312 I kl_divergence: computing over 44 chunks, n_ctx=2048, batch_size=2048, n_seq=1
|
| 4 |
+
0.05.780.560 I kl_divergence: 3.65 seconds per pass - ETA 2.67 minutes
|
| 5 |
+
|
| 6 |
+
chunk PPL ln(PPL(Q)/PPL(base)) KL Divergence Δp RMS Same top p
|
| 7 |
+
1 42.7479 ± 4.1534 0.08668 ± 0.03082 0.29717 ± 0.01140 11.898 ± 0.638 % 70.479 ± 1.427 %
|
| 8 |
+
2 47.6359 ± 3.3896 0.09627 ± 0.02228 0.29492 ± 0.00839 11.628 ± 0.483 % 71.065 ± 1.003 %
|
| 9 |
+
3 33.5703 ± 1.9598 0.08225 ± 0.01670 0.25226 ± 0.00658 10.990 ± 0.407 % 74.389 ± 0.788 %
|
| 10 |
+
4 22.5348 ± 1.1245 0.06065 ± 0.01429 0.23519 ± 0.00605 11.056 ± 0.355 % 76.931 ± 0.659 %
|
| 11 |
+
5 17.7061 ± 0.7722 0.05625 ± 0.01228 0.20981 ± 0.00535 10.936 ± 0.328 % 79.277 ± 0.567 %
|
| 12 |
+
6 20.3186 ± 0.8182 0.07232 ± 0.01141 0.23402 ± 0.00516 11.256 ± 0.293 % 77.566 ± 0.532 %
|
| 13 |
+
7 23.0720 ± 0.8697 0.07592 ± 0.01066 0.24514 ± 0.00478 11.324 ± 0.269 % 76.288 ± 0.503 %
|
| 14 |
+
8 25.5785 ± 0.9040 0.08081 ± 0.01008 0.25942 ± 0.00458 11.354 ± 0.247 % 75.134 ± 0.478 %
|
| 15 |
+
9 22.3535 ± 0.7405 0.07635 ± 0.00952 0.25430 ± 0.00453 11.657 ± 0.254 % 76.116 ± 0.444 %
|
| 16 |
+
10 25.1425 ± 0.8005 0.07891 ± 0.00903 0.25970 ± 0.00421 11.540 ± 0.238 % 75.034 ± 0.428 %
|
| 17 |
+
11 22.3257 ± 0.6730 0.07263 ± 0.00862 0.25585 ± 0.00412 11.767 ± 0.232 % 75.873 ± 0.403 %
|
| 18 |
+
12 21.6625 ± 0.6205 0.08121 ± 0.00819 0.25117 ± 0.00386 11.785 ± 0.217 % 76.303 ± 0.384 %
|
| 19 |
+
13 21.7052 ± 0.5958 0.09052 ± 0.00794 0.26004 ± 0.00373 12.010 ± 0.205 % 75.652 ± 0.372 %
|
| 20 |
+
14 23.0156 ± 0.6114 0.09185 ± 0.00785 0.27259 ± 0.00379 12.104 ± 0.197 % 74.843 ± 0.363 %
|
| 21 |
+
15 22.1689 ± 0.5661 0.09114 ± 0.00759 0.27006 ± 0.00361 12.263 ± 0.190 % 74.963 ± 0.350 %
|
| 22 |
+
16 20.4489 ± 0.5031 0.08678 ± 0.00731 0.26437 ± 0.00360 12.378 ± 0.189 % 75.611 ± 0.336 %
|
| 23 |
+
17 20.2855 ± 0.4812 0.08323 ± 0.00710 0.26529 ± 0.00345 12.346 ± 0.180 % 75.585 ± 0.326 %
|
| 24 |
+
18 20.9329 ± 0.4821 0.08625 ± 0.00683 0.26274 ± 0.00329 12.240 ± 0.173 % 75.377 ± 0.317 %
|
| 25 |
+
19 21.2465 ± 0.4762 0.09307 ± 0.00670 0.26558 ± 0.00319 12.262 ± 0.167 % 75.217 ± 0.310 %
|
| 26 |
+
20 21.7338 ± 0.4741 0.09425 ± 0.00656 0.26955 ± 0.00311 12.232 ± 0.161 % 74.961 ± 0.303 %
|
| 27 |
+
21 22.1211 ± 0.4723 0.10436 ± 0.00643 0.27138 ± 0.00303 12.162 ± 0.156 % 74.910 ± 0.296 %
|
| 28 |
+
22 21.2445 ± 0.4404 0.10191 ± 0.00624 0.26645 ± 0.00300 12.110 ± 0.153 % 75.260 ± 0.288 %
|
| 29 |
+
23 20.0829 ± 0.4053 0.09935 ± 0.00604 0.26110 ± 0.00291 12.072 ± 0.150 % 75.643 ± 0.280 %
|
| 30 |
+
24 20.2131 ± 0.4001 0.09806 ± 0.00595 0.26261 ± 0.00286 12.108 ± 0.146 % 75.497 ± 0.274 %
|
| 31 |
+
25 19.4393 ± 0.3756 0.09633 ± 0.00575 0.25676 ± 0.00277 12.003 ± 0.143 % 75.937 ± 0.267 %
|
| 32 |
+
26 19.7562 ± 0.3738 0.09608 ± 0.00561 0.25453 ± 0.00267 11.947 ± 0.139 % 75.874 ± 0.262 %
|
| 33 |
+
27 18.7920 ± 0.3474 0.09399 ± 0.00549 0.25243 ± 0.00264 11.995 ± 0.137 % 76.178 ± 0.256 %
|
| 34 |
+
28 19.0032 ± 0.3443 0.09289 ± 0.00537 0.25198 ± 0.00257 11.975 ± 0.134 % 76.110 ± 0.252 %
|
| 35 |
+
29 19.3639 ± 0.3458 0.09432 ± 0.00529 0.25422 ± 0.00253 12.049 ± 0.132 % 75.953 ± 0.248 %
|
| 36 |
+
30 19.8213 ± 0.3481 0.09523 ± 0.00523 0.25674 ± 0.00250 12.017 ± 0.129 % 75.751 ± 0.245 %
|
| 37 |
+
31 19.6233 ± 0.3376 0.09455 ± 0.00513 0.25581 ± 0.00245 12.030 ± 0.127 % 75.707 ± 0.241 %
|
| 38 |
+
32 19.9074 ± 0.3375 0.12328 ± 0.00546 0.29033 ± 0.00299 13.298 ± 0.145 % 74.875 ± 0.240 %
|
| 39 |
+
33 20.1685 ± 0.3361 0.12232 ± 0.00537 0.28906 ± 0.00292 13.241 ± 0.142 % 74.866 ± 0.236 %
|
| 40 |
+
34 20.2095 ± 0.3309 0.12204 ± 0.00526 0.28718 ± 0.00285 13.203 ± 0.139 % 74.780 ± 0.233 %
|
| 41 |
+
35 20.4523 ± 0.3306 0.12090 ± 0.00520 0.28946 �� 0.00280 13.230 ± 0.136 % 74.629 ± 0.230 %
|
| 42 |
+
36 20.7814 ± 0.3325 0.11060 ± 0.00550 0.33851 ± 0.00368 13.804 ± 0.138 % 73.727 ± 0.229 %
|
| 43 |
+
37 21.0618 ± 0.3323 0.11193 ± 0.00539 0.33515 ± 0.00359 13.719 ± 0.136 % 73.731 ± 0.226 %
|
| 44 |
+
38 21.5154 ± 0.3356 0.11177 ± 0.00530 0.33381 ± 0.00351 13.659 ± 0.133 % 73.661 ± 0.223 %
|
| 45 |
+
39 21.8036 ± 0.3358 0.11326 ± 0.00520 0.33132 ± 0.00343 13.582 ± 0.131 % 73.697 ± 0.220 %
|
| 46 |
+
40 21.9536 ± 0.3338 0.11425 ± 0.00512 0.32989 ± 0.00335 13.535 ± 0.129 % 73.678 ± 0.218 %
|
| 47 |
+
41 22.2466 ± 0.3346 0.11551 ± 0.00504 0.32881 ± 0.00328 13.482 ± 0.127 % 73.619 ± 0.215 %
|
| 48 |
+
42 22.5787 ± 0.3362 0.11744 ± 0.00498 0.32853 ± 0.00322 13.436 ± 0.125 % 73.542 ± 0.213 %
|
| 49 |
+
43 22.9255 ± 0.3376 0.11785 ± 0.00491 0.32810 ± 0.00317 13.396 ± 0.124 % 73.505 ± 0.210 %
|
| 50 |
+
44 22.4858 ± 0.3263 0.11597 ± 0.00481 0.32229 ± 0.00311 13.278 ± 0.122 % 73.760 ± 0.207 %
|
| 51 |
+
|
| 52 |
+
====== Perplexity statistics ======
|
| 53 |
+
Mean PPL(Q) : 22.485774 ± 0.326277
|
| 54 |
+
Mean PPL(base) : 20.023532 ± 0.285215
|
| 55 |
+
Cor(ln(PPL(Q)), ln(PPL(base))): 94.42%
|
| 56 |
+
Mean ln(PPL(Q)/PPL(base)) : 0.115975 ± 0.004810
|
| 57 |
+
Mean PPL(Q)/PPL(base) : 1.122967 ± 0.005402
|
| 58 |
+
Mean PPL(Q)-PPL(base) : 2.462243 ± 0.109870
|
| 59 |
+
|
| 60 |
+
====== KL divergence statistics ======
|
| 61 |
+
Mean KLD: 0.322294 ± 0.003108
|
| 62 |
+
Maximum KLD: 15.541307
|
| 63 |
+
99.9% KLD: 7.824029
|
| 64 |
+
99.0% KLD: 3.186437
|
| 65 |
+
95.0% KLD: 1.145174
|
| 66 |
+
90.0% KLD: 0.691000
|
| 67 |
+
Median KLD: 0.148998
|
| 68 |
+
10.0% KLD: 0.009049
|
| 69 |
+
5.0% KLD: 0.001746
|
| 70 |
+
1.0% KLD: 0.000052
|
| 71 |
+
0.1% KLD: 0.000001
|
| 72 |
+
Minimum KLD: -0.000018
|
| 73 |
+
|
| 74 |
+
====== Token probability statistics ======
|
| 75 |
+
Mean Δp: -1.378 ± 0.062 %
|
| 76 |
+
Maximum Δp: 99.759%
|
| 77 |
+
99.9% Δp: 78.834%
|
| 78 |
+
99.0% Δp: 36.834%
|
| 79 |
+
95.0% Δp: 15.041%
|
| 80 |
+
90.0% Δp: 7.773%
|
| 81 |
+
75.0% Δp: 0.964%
|
| 82 |
+
Median Δp: -0.017%
|
| 83 |
+
25.0% Δp: -2.497%
|
| 84 |
+
10.0% Δp: -12.388%
|
| 85 |
+
5.0% Δp: -22.110%
|
| 86 |
+
1.0% Δp: -50.488%
|
| 87 |
+
0.1% Δp: -92.557%
|
| 88 |
+
Minimum Δp: -99.918%
|
| 89 |
+
RMS Δp : 13.278 ± 0.122 %
|
| 90 |
+
Same top p: 73.760 ± 0.207 %
|
| 91 |
+
|
reports/kld2048-Q4_K_M.txt
ADDED
|
@@ -0,0 +1,91 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
0.02.444.999 I
|
| 2 |
+
0.02.445.022 I system_info: n_threads = 12 (n_threads_batch = 12) / 16 | MTL : EMBED_LIBRARY = 1 | CPU : NEON = 1 | ARM_FMA = 1 | FP16_VA = 1 | MATMUL_INT8 = 1 | DOTPROD = 1 | SME = 1 | ACCELERATE = 1 | OPENMP = 1 | REPACK = 1 |
|
| 3 |
+
0.02.458.394 I kl_divergence: computing over 44 chunks, n_ctx=2048, batch_size=2048, n_seq=1
|
| 4 |
+
0.06.572.750 I kl_divergence: 4.11 seconds per pass - ETA 3.02 minutes
|
| 5 |
+
|
| 6 |
+
chunk PPL ln(PPL(Q)/PPL(base)) KL Divergence Δp RMS Same top p
|
| 7 |
+
1 41.2795 ± 4.1398 0.05173 ± 0.02221 0.14531 ± 0.00574 8.854 ± 0.552 % 77.126 ± 1.314 %
|
| 8 |
+
2 45.0236 ± 3.2155 0.03987 ± 0.01546 0.14469 ± 0.00415 8.361 ± 0.366 % 78.250 ± 0.912 %
|
| 9 |
+
3 31.5781 ± 1.8364 0.02107 ± 0.01123 0.11655 ± 0.00343 7.819 ± 0.360 % 82.079 ± 0.692 %
|
| 10 |
+
4 21.6469 ± 1.0758 0.02045 ± 0.00935 0.10766 ± 0.00419 7.739 ± 0.332 % 83.993 ± 0.573 %
|
| 11 |
+
5 16.9931 ± 0.7377 0.01515 ± 0.00792 0.09377 ± 0.00344 7.450 ± 0.295 % 85.885 ± 0.487 %
|
| 12 |
+
6 19.4869 ± 0.7820 0.03052 ± 0.00775 0.10659 ± 0.00321 7.721 ± 0.261 % 84.653 ± 0.460 %
|
| 13 |
+
7 22.6632 ± 0.8578 0.05804 ± 0.00758 0.11565 ± 0.00297 7.946 ± 0.239 % 83.550 ± 0.438 %
|
| 14 |
+
8 20.9572 ± 0.7326 -0.11846 ± 0.01231 0.39262 ± 0.01300 14.625 ± 0.373 % 78.434 ± 0.455 %
|
| 15 |
+
9 18.5839 ± 0.6074 -0.10834 ± 0.01109 0.35627 ± 0.01162 14.009 ± 0.348 % 79.700 ± 0.419 %
|
| 16 |
+
10 21.2826 ± 0.6718 -0.08776 ± 0.01025 0.33616 ± 0.01050 13.492 ± 0.328 % 79.413 ± 0.400 %
|
| 17 |
+
11 19.1917 ± 0.5743 -0.07863 ± 0.00945 0.31249 ± 0.00958 13.037 ± 0.309 % 80.450 ± 0.374 %
|
| 18 |
+
12 19.5849 ± 0.5619 -0.01962 ± 0.00936 0.32527 ± 0.00897 13.301 ± 0.289 % 79.822 ± 0.362 %
|
| 19 |
+
13 19.5401 ± 0.5360 -0.01456 ± 0.00876 0.31061 ± 0.00831 13.029 ± 0.274 % 79.946 ± 0.347 %
|
| 20 |
+
14 20.8731 ± 0.5551 -0.00586 ± 0.00827 0.30007 ± 0.00774 12.756 ± 0.261 % 79.807 ± 0.335 %
|
| 21 |
+
15 20.1689 ± 0.5160 -0.00341 ± 0.00783 0.28704 ± 0.00724 12.577 ± 0.249 % 80.137 ± 0.322 %
|
| 22 |
+
16 18.6731 ± 0.4595 -0.00407 ± 0.00741 0.27279 ± 0.00682 12.324 ± 0.240 % 80.816 ± 0.308 %
|
| 23 |
+
17 18.6061 ± 0.4410 -0.00319 ± 0.00707 0.26420 ± 0.00643 12.100 ± 0.231 % 80.823 ± 0.299 %
|
| 24 |
+
18 19.1916 ± 0.4411 -0.00060 ± 0.00674 0.25503 ± 0.00608 11.886 ± 0.223 % 80.884 ± 0.290 %
|
| 25 |
+
19 19.4546 ± 0.4344 0.00496 ± 0.00649 0.24983 ± 0.00577 11.775 ± 0.214 % 80.794 ± 0.283 %
|
| 26 |
+
20 19.9924 ± 0.4346 0.01074 ± 0.00625 0.24438 ± 0.00549 11.600 ± 0.207 % 80.689 ± 0.276 %
|
| 27 |
+
21 20.2596 ± 0.4294 0.01645 ± 0.00602 0.23860 ± 0.00524 11.452 ± 0.200 % 80.752 ± 0.269 %
|
| 28 |
+
22 19.5131 ± 0.4014 0.01690 ± 0.00577 0.23014 ± 0.00501 11.276 ± 0.195 % 81.112 ± 0.261 %
|
| 29 |
+
23 18.4590 ± 0.3693 0.01504 ± 0.00555 0.22207 ± 0.00480 11.100 ± 0.190 % 81.550 ± 0.253 %
|
| 30 |
+
24 18.6444 ± 0.3663 0.01727 ± 0.00542 0.21940 ± 0.00462 11.020 ± 0.184 % 81.496 ± 0.248 %
|
| 31 |
+
25 17.9511 ± 0.3440 0.01668 ± 0.00522 0.21212 ± 0.00444 10.845 ± 0.180 % 81.896 ± 0.241 %
|
| 32 |
+
26 18.2686 ± 0.3430 0.01779 ± 0.00506 0.20778 ± 0.00428 10.719 ± 0.175 % 81.908 ± 0.236 %
|
| 33 |
+
27 17.4065 ± 0.3197 0.01740 ± 0.00493 0.20349 ± 0.00413 10.692 ± 0.171 % 82.206 ± 0.230 %
|
| 34 |
+
28 17.6385 ± 0.3178 0.01837 ± 0.00479 0.20009 ± 0.00399 10.591 ± 0.167 % 82.216 ± 0.226 %
|
| 35 |
+
29 18.0042 ± 0.3199 0.02152 ± 0.00468 0.19849 ± 0.00386 10.544 ± 0.162 % 82.149 ± 0.222 %
|
| 36 |
+
30 18.4178 ± 0.3215 0.02179 ± 0.00459 0.19651 ± 0.00374 10.470 ± 0.159 % 82.030 ± 0.219 %
|
| 37 |
+
31 18.2709 ± 0.3124 0.02314 ± 0.00447 0.19322 ± 0.00362 10.412 ± 0.155 % 82.121 ± 0.215 %
|
| 38 |
+
32 17.8423 ± 0.2996 0.01376 ± 0.00451 0.19735 ± 0.00358 10.703 ± 0.155 % 82.005 ± 0.212 %
|
| 39 |
+
33 18.1069 ± 0.2990 0.01449 ± 0.00441 0.19455 ± 0.00348 10.613 ± 0.152 % 82.014 ± 0.209 %
|
| 40 |
+
34 18.1601 ± 0.2946 0.01511 ± 0.00432 0.19198 ± 0.00338 10.562 ± 0.148 % 82.014 ± 0.206 %
|
| 41 |
+
35 18.4193 ± 0.2951 0.01620 ± 0.00424 0.19046 �� 0.00329 10.519 ± 0.145 % 81.997 ± 0.203 %
|
| 42 |
+
36 18.9312 ± 0.3005 0.01736 ± 0.00433 0.20369 ± 0.00333 10.673 ± 0.142 % 81.435 ± 0.203 %
|
| 43 |
+
37 19.1886 ± 0.3005 0.01879 ± 0.00423 0.20058 ± 0.00325 10.580 ± 0.140 % 81.549 ± 0.199 %
|
| 44 |
+
38 19.5951 ± 0.3033 0.01828 ± 0.00415 0.19855 ± 0.00317 10.517 ± 0.138 % 81.535 ± 0.197 %
|
| 45 |
+
39 19.8455 ± 0.3031 0.01916 ± 0.00406 0.19609 ± 0.00309 10.454 ± 0.135 % 81.542 ± 0.194 %
|
| 46 |
+
40 19.9566 ± 0.3008 0.01888 ± 0.00399 0.19379 ± 0.00301 10.382 ± 0.133 % 81.569 ± 0.192 %
|
| 47 |
+
41 20.2158 ± 0.3012 0.01978 ± 0.00392 0.19213 ± 0.00294 10.323 ± 0.131 % 81.480 ± 0.190 %
|
| 48 |
+
42 20.5107 ± 0.3025 0.02138 ± 0.00386 0.19120 ± 0.00288 10.284 ± 0.128 % 81.434 ± 0.188 %
|
| 49 |
+
43 20.8351 ± 0.3039 0.02224 ± 0.00379 0.19001 ± 0.00282 10.229 ± 0.126 % 81.395 ± 0.186 %
|
| 50 |
+
44 20.4607 ± 0.2940 0.02160 ± 0.00371 0.18611 ± 0.00276 10.126 ± 0.125 % 81.632 ± 0.183 %
|
| 51 |
+
|
| 52 |
+
====== Perplexity statistics ======
|
| 53 |
+
Mean PPL(Q) : 20.460742 ± 0.294034
|
| 54 |
+
Mean PPL(base) : 20.023532 ± 0.285215
|
| 55 |
+
Cor(ln(PPL(Q)), ln(PPL(base))): 96.64%
|
| 56 |
+
Mean ln(PPL(Q)/PPL(base)) : 0.021600 ± 0.003713
|
| 57 |
+
Mean PPL(Q)/PPL(base) : 1.021835 ± 0.003794
|
| 58 |
+
Mean PPL(Q)-PPL(base) : 0.437210 ± 0.075617
|
| 59 |
+
|
| 60 |
+
====== KL divergence statistics ======
|
| 61 |
+
Mean KLD: 0.186110 ± 0.002757
|
| 62 |
+
Maximum KLD: 18.745331
|
| 63 |
+
99.9% KLD: 8.276571
|
| 64 |
+
99.0% KLD: 2.376208
|
| 65 |
+
95.0% KLD: 0.638440
|
| 66 |
+
90.0% KLD: 0.343436
|
| 67 |
+
Median KLD: 0.062474
|
| 68 |
+
10.0% KLD: 0.003218
|
| 69 |
+
5.0% KLD: 0.000575
|
| 70 |
+
1.0% KLD: 0.000016
|
| 71 |
+
0.1% KLD: 0.000000
|
| 72 |
+
Minimum KLD: -0.000008
|
| 73 |
+
|
| 74 |
+
====== Token probability statistics ======
|
| 75 |
+
Mean Δp: -0.062 ± 0.048 %
|
| 76 |
+
Maximum Δp: 99.658%
|
| 77 |
+
99.9% Δp: 90.783%
|
| 78 |
+
99.0% Δp: 34.222%
|
| 79 |
+
95.0% Δp: 11.752%
|
| 80 |
+
90.0% Δp: 5.987%
|
| 81 |
+
75.0% Δp: 0.861%
|
| 82 |
+
Median Δp: -0.003%
|
| 83 |
+
25.0% Δp: -1.287%
|
| 84 |
+
10.0% Δp: -7.002%
|
| 85 |
+
5.0% Δp: -12.914%
|
| 86 |
+
1.0% Δp: -30.396%
|
| 87 |
+
0.1% Δp: -62.846%
|
| 88 |
+
Minimum Δp: -98.875%
|
| 89 |
+
RMS Δp : 10.126 ± 0.125 %
|
| 90 |
+
Same top p: 81.632 ± 0.183 %
|
| 91 |
+
|
reports/kld2048-Q5_K_M.txt
ADDED
|
@@ -0,0 +1,91 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
0.02.768.224 I
|
| 2 |
+
0.02.768.244 I system_info: n_threads = 12 (n_threads_batch = 12) / 16 | MTL : EMBED_LIBRARY = 1 | CPU : NEON = 1 | ARM_FMA = 1 | FP16_VA = 1 | MATMUL_INT8 = 1 | DOTPROD = 1 | SME = 1 | ACCELERATE = 1 | OPENMP = 1 | REPACK = 1 |
|
| 3 |
+
0.02.782.686 I kl_divergence: computing over 44 chunks, n_ctx=2048, batch_size=2048, n_seq=1
|
| 4 |
+
0.07.371.405 I kl_divergence: 4.59 seconds per pass - ETA 3.35 minutes
|
| 5 |
+
|
| 6 |
+
chunk PPL ln(PPL(Q)/PPL(base)) KL Divergence Δp RMS Same top p
|
| 7 |
+
1 38.8757 ± 3.9445 -0.00827 ± 0.01320 0.05545 ± 0.00215 5.215 ± 0.340 % 85.337 ± 1.107 %
|
| 8 |
+
2 42.6337 ± 3.0419 -0.01467 ± 0.01052 0.06207 ± 0.00192 5.466 ± 0.256 % 85.679 ± 0.775 %
|
| 9 |
+
3 30.5964 ± 1.7826 -0.01051 ± 0.00743 0.04750 ± 0.00140 4.847 ± 0.208 % 88.335 ± 0.580 %
|
| 10 |
+
4 20.9749 ± 1.0406 -0.01108 ± 0.00618 0.04456 ± 0.00308 4.891 ± 0.299 % 89.565 ± 0.478 %
|
| 11 |
+
5 16.5848 ± 0.7193 -0.00917 ± 0.00518 0.03800 ± 0.00249 4.607 ± 0.262 % 90.811 ± 0.404 %
|
| 12 |
+
6 18.7215 ± 0.7472 -0.00954 ± 0.00509 0.04528 ± 0.00224 5.137 ± 0.252 % 89.834 ± 0.386 %
|
| 13 |
+
7 21.3836 ± 0.8000 -0.00008 ± 0.00490 0.04909 ± 0.00199 5.227 ± 0.226 % 89.345 ± 0.365 %
|
| 14 |
+
8 23.6569 ± 0.8287 0.00271 ± 0.00471 0.05155 ± 0.00178 5.248 ± 0.202 % 88.685 ± 0.350 %
|
| 15 |
+
9 20.7981 ± 0.6838 0.00422 ± 0.00433 0.04807 ± 0.00160 5.076 ± 0.187 % 89.443 ± 0.320 %
|
| 16 |
+
10 23.1873 ± 0.7314 -0.00204 ± 0.00410 0.05020 ± 0.00148 5.046 ± 0.173 % 88.993 ± 0.309 %
|
| 17 |
+
11 20.7451 ± 0.6213 -0.00079 ± 0.00383 0.04784 ± 0.00136 4.998 ± 0.161 % 89.469 ± 0.289 %
|
| 18 |
+
12 19.8598 ± 0.5643 -0.00567 ± 0.00364 0.04770 ± 0.00128 5.168 ± 0.155 % 89.492 ± 0.277 %
|
| 19 |
+
13 19.7849 ± 0.5377 -0.00211 ± 0.00350 0.04857 ± 0.00120 5.245 ± 0.146 % 89.240 ± 0.269 %
|
| 20 |
+
14 20.9128 ± 0.5502 -0.00396 ± 0.00342 0.04971 ± 0.00113 5.239 ± 0.137 % 88.975 ± 0.262 %
|
| 21 |
+
15 20.1519 ± 0.5098 -0.00425 ± 0.00332 0.04973 ± 0.00107 5.337 ± 0.131 % 89.078 ± 0.252 %
|
| 22 |
+
16 18.6664 ± 0.4549 -0.00443 ± 0.00315 0.04782 ± 0.00102 5.274 ± 0.126 % 89.479 ± 0.240 %
|
| 23 |
+
17 18.5783 ± 0.4353 -0.00469 ± 0.00302 0.04728 ± 0.00097 5.242 ± 0.120 % 89.397 ± 0.233 %
|
| 24 |
+
18 19.1103 ± 0.4339 -0.00485 ± 0.00290 0.04656 ± 0.00092 5.224 ± 0.116 % 89.394 ± 0.227 %
|
| 25 |
+
19 19.3085 ± 0.4256 -0.00258 ± 0.00282 0.04711 ± 0.00088 5.253 ± 0.111 % 89.299 ± 0.222 %
|
| 26 |
+
20 19.7024 ± 0.4220 -0.00387 ± 0.00276 0.04716 ± 0.00084 5.231 ± 0.107 % 89.262 ± 0.216 %
|
| 27 |
+
21 19.8591 ± 0.4144 -0.00351 ± 0.00268 0.04710 ± 0.00082 5.209 ± 0.103 % 89.308 ± 0.211 %
|
| 28 |
+
22 19.1254 ± 0.3875 -0.00317 ± 0.00257 0.04570 ± 0.00078 5.159 ± 0.101 % 89.576 ± 0.204 %
|
| 29 |
+
23 18.1151 ± 0.3572 -0.00377 ± 0.00248 0.04447 ± 0.00077 5.115 ± 0.100 % 89.885 ± 0.197 %
|
| 30 |
+
24 18.2710 ± 0.3538 -0.00296 ± 0.00244 0.04467 ± 0.00075 5.161 ± 0.098 % 89.781 ± 0.193 %
|
| 31 |
+
25 17.6049 ± 0.3328 -0.00279 ± 0.00236 0.04338 ± 0.00072 5.088 ± 0.095 % 89.975 ± 0.188 %
|
| 32 |
+
26 17.8942 ± 0.3316 -0.00291 ± 0.00231 0.04337 ± 0.00069 5.059 ± 0.092 % 89.943 ± 0.184 %
|
| 33 |
+
27 17.0782 ± 0.3097 -0.00165 ± 0.00229 0.04387 ± 0.00071 5.213 ± 0.095 % 90.000 ± 0.181 %
|
| 34 |
+
28 17.2875 ± 0.3074 -0.00173 ± 0.00224 0.04372 ± 0.00068 5.201 ± 0.093 % 89.994 ± 0.177 %
|
| 35 |
+
29 17.5791 ± 0.3081 -0.00238 ± 0.00222 0.04433 ± 0.00067 5.221 ± 0.090 % 89.878 ± 0.175 %
|
| 36 |
+
30 17.9585 ± 0.3092 -0.00347 ± 0.00219 0.04487 ± 0.00066 5.215 ± 0.088 % 89.713 ± 0.173 %
|
| 37 |
+
31 17.8034 ± 0.3002 -0.00278 ± 0.00215 0.04463 ± 0.00064 5.224 ± 0.087 % 89.749 ± 0.170 %
|
| 38 |
+
32 17.3903 ± 0.2882 -0.01190 ± 0.00246 0.05683 ± 0.00117 6.108 ± 0.117 % 89.260 ± 0.171 %
|
| 39 |
+
33 17.6195 ± 0.2873 -0.01280 ± 0.00241 0.05650 ± 0.00114 6.078 ± 0.115 % 89.188 ± 0.169 %
|
| 40 |
+
34 17.6545 ± 0.2828 -0.01312 ± 0.00236 0.05607 ± 0.00110 6.052 ± 0.112 % 89.098 ± 0.167 %
|
| 41 |
+
35 17.8969 ± 0.2831 -0.01257 ± 0.00233 0.05668 �� 0.00108 6.079 ± 0.110 % 88.996 ± 0.165 %
|
| 42 |
+
36 18.3681 ± 0.2879 -0.01284 ± 0.00233 0.05805 ± 0.00107 6.104 ± 0.107 % 88.816 ± 0.164 %
|
| 43 |
+
37 18.5831 ± 0.2872 -0.01327 ± 0.00228 0.05735 ± 0.00104 6.053 ± 0.105 % 88.888 ± 0.162 %
|
| 44 |
+
38 18.9709 ± 0.2897 -0.01409 ± 0.00224 0.05711 ± 0.00102 6.017 ± 0.103 % 88.797 ± 0.160 %
|
| 45 |
+
39 19.1935 ± 0.2891 -0.01424 ± 0.00220 0.05668 ± 0.00099 5.984 ± 0.101 % 88.794 ± 0.158 %
|
| 46 |
+
40 19.2943 ± 0.2866 -0.01487 ± 0.00216 0.05641 ± 0.00097 5.952 ± 0.099 % 88.761 ± 0.156 %
|
| 47 |
+
41 19.5044 ± 0.2863 -0.01604 ± 0.00213 0.05616 ± 0.00095 5.919 ± 0.097 % 88.713 ± 0.155 %
|
| 48 |
+
42 19.7541 ± 0.2869 -0.01621 ± 0.00210 0.05625 ± 0.00093 5.898 ± 0.096 % 88.658 ± 0.153 %
|
| 49 |
+
43 20.0280 ± 0.2875 -0.01727 ± 0.00207 0.05609 ± 0.00091 5.880 ± 0.094 % 88.631 ± 0.151 %
|
| 50 |
+
44 19.6851 ± 0.2785 -0.01705 ± 0.00203 0.05497 ± 0.00089 5.822 ± 0.093 % 88.785 ± 0.149 %
|
| 51 |
+
|
| 52 |
+
====== Perplexity statistics ======
|
| 53 |
+
Mean PPL(Q) : 19.685105 ± 0.278469
|
| 54 |
+
Mean PPL(base) : 20.023532 ± 0.285215
|
| 55 |
+
Cor(ln(PPL(Q)), ln(PPL(base))): 98.98%
|
| 56 |
+
Mean ln(PPL(Q)/PPL(base)) : -0.017046 ± 0.002029
|
| 57 |
+
Mean PPL(Q)/PPL(base) : 0.983099 ± 0.001994
|
| 58 |
+
Mean PPL(Q)-PPL(base) : -0.338427 ± 0.040788
|
| 59 |
+
|
| 60 |
+
====== KL divergence statistics ======
|
| 61 |
+
Mean KLD: 0.054971 ± 0.000887
|
| 62 |
+
Maximum KLD: 15.123309
|
| 63 |
+
99.9% KLD: 2.112881
|
| 64 |
+
99.0% KLD: 0.528482
|
| 65 |
+
95.0% KLD: 0.179696
|
| 66 |
+
90.0% KLD: 0.109617
|
| 67 |
+
Median KLD: 0.022553
|
| 68 |
+
10.0% KLD: 0.001180
|
| 69 |
+
5.0% KLD: 0.000204
|
| 70 |
+
1.0% KLD: 0.000005
|
| 71 |
+
0.1% KLD: -0.000001
|
| 72 |
+
Minimum KLD: -0.000105
|
| 73 |
+
|
| 74 |
+
====== Token probability statistics ======
|
| 75 |
+
Mean Δp: -0.061 ± 0.027 %
|
| 76 |
+
Maximum Δp: 99.631%
|
| 77 |
+
99.9% Δp: 50.258%
|
| 78 |
+
99.0% Δp: 17.778%
|
| 79 |
+
95.0% Δp: 6.754%
|
| 80 |
+
90.0% Δp: 3.472%
|
| 81 |
+
75.0% Δp: 0.548%
|
| 82 |
+
Median Δp: 0.000%
|
| 83 |
+
25.0% Δp: -0.714%
|
| 84 |
+
10.0% Δp: -3.974%
|
| 85 |
+
5.0% Δp: -7.296%
|
| 86 |
+
1.0% Δp: -17.461%
|
| 87 |
+
0.1% Δp: -40.127%
|
| 88 |
+
Minimum Δp: -91.966%
|
| 89 |
+
RMS Δp : 5.822 ± 0.093 %
|
| 90 |
+
Same top p: 88.785 ± 0.149 %
|
| 91 |
+
|
reports/kld2048-Q6_K.txt
ADDED
|
@@ -0,0 +1,91 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
0.00.753.072 I
|
| 2 |
+
0.00.753.091 I system_info: n_threads = 12 (n_threads_batch = 12) / 16 | MTL : EMBED_LIBRARY = 1 | CPU : NEON = 1 | ARM_FMA = 1 | FP16_VA = 1 | MATMUL_INT8 = 1 | DOTPROD = 1 | SME = 1 | ACCELERATE = 1 | OPENMP = 1 | REPACK = 1 |
|
| 3 |
+
0.00.766.732 I kl_divergence: computing over 44 chunks, n_ctx=2048, batch_size=2048, n_seq=1
|
| 4 |
+
0.05.072.707 I kl_divergence: 4.31 seconds per pass - ETA 3.15 minutes
|
| 5 |
+
|
| 6 |
+
chunk PPL ln(PPL(Q)/PPL(base)) KL Divergence Δp RMS Same top p
|
| 7 |
+
1 40.0041 ± 4.0847 0.02035 ± 0.01294 0.04754 ± 0.00217 5.017 ± 0.394 % 86.315 ± 1.075 %
|
| 8 |
+
2 43.1959 ± 3.1069 -0.00157 ± 0.00981 0.05240 ± 0.00169 5.029 ± 0.265 % 86.510 ± 0.755 %
|
| 9 |
+
3 30.7769 ± 1.8008 -0.00463 ± 0.00686 0.03930 ± 0.00120 4.418 ± 0.209 % 89.215 ± 0.560 %
|
| 10 |
+
4 21.0446 ± 1.0486 -0.00777 ± 0.00564 0.03561 ± 0.00234 4.555 ± 0.312 % 90.616 ± 0.456 %
|
| 11 |
+
5 16.6295 ± 0.7240 -0.00648 ± 0.00462 0.02974 ± 0.00188 4.205 ± 0.272 % 91.984 ± 0.380 %
|
| 12 |
+
6 18.8923 ± 0.7585 -0.00047 ± 0.00447 0.03427 ± 0.00164 4.595 ± 0.233 % 91.137 ± 0.363 %
|
| 13 |
+
7 21.4123 ± 0.8052 0.00126 ± 0.00436 0.03803 ± 0.00150 4.725 ± 0.210 % 90.476 ± 0.347 %
|
| 14 |
+
8 23.6670 ± 0.8334 0.00314 ± 0.00425 0.04068 ± 0.00135 4.725 ± 0.186 % 89.944 ± 0.332 %
|
| 15 |
+
9 20.7545 ± 0.6848 0.00213 ± 0.00387 0.03769 ± 0.00121 4.555 ± 0.173 % 90.518 ± 0.305 %
|
| 16 |
+
10 23.1620 ± 0.7331 -0.00314 ± 0.00368 0.03983 ± 0.00113 4.540 ± 0.159 % 89.873 ± 0.298 %
|
| 17 |
+
11 20.6827 ± 0.6208 -0.00380 ± 0.00342 0.03767 ± 0.00103 4.448 ± 0.148 % 90.483 ± 0.277 %
|
| 18 |
+
12 20.0255 ± 0.5712 0.00263 ± 0.00340 0.04175 ± 0.00102 4.825 ± 0.146 % 90.135 ± 0.269 %
|
| 19 |
+
13 19.9151 ± 0.5433 0.00445 ± 0.00324 0.04214 ± 0.00096 4.911 ± 0.138 % 90.044 ± 0.260 %
|
| 20 |
+
14 20.9747 ± 0.5539 -0.00100 ± 0.00320 0.04403 ± 0.00093 4.946 ± 0.129 % 89.548 ± 0.256 %
|
| 21 |
+
15 20.2122 ± 0.5131 -0.00127 ± 0.00310 0.04383 ± 0.00088 5.037 ± 0.123 % 89.606 ± 0.246 %
|
| 22 |
+
16 18.7312 ± 0.4580 -0.00096 ± 0.00293 0.04176 ± 0.00083 4.944 ± 0.118 % 89.968 ± 0.235 %
|
| 23 |
+
17 18.6427 ± 0.4385 -0.00122 ± 0.00282 0.04110 ± 0.00079 4.885 ± 0.113 % 89.960 ± 0.228 %
|
| 24 |
+
18 19.1412 ± 0.4364 -0.00323 ± 0.00270 0.04022 ± 0.00075 4.838 ± 0.108 % 90.089 ± 0.220 %
|
| 25 |
+
19 19.3194 ± 0.4275 -0.00201 ± 0.00263 0.04049 ± 0.00072 4.851 ± 0.104 % 89.993 ± 0.215 %
|
| 26 |
+
20 19.7105 ± 0.4240 -0.00346 ± 0.00256 0.04038 ± 0.00069 4.818 ± 0.100 % 89.922 ± 0.210 %
|
| 27 |
+
21 19.8347 ± 0.4156 -0.00474 ± 0.00247 0.03983 ± 0.00066 4.776 ± 0.097 % 90.029 ± 0.204 %
|
| 28 |
+
22 19.1081 ± 0.3888 -0.00407 ± 0.00237 0.03851 ± 0.00063 4.704 ± 0.094 % 90.287 ± 0.197 %
|
| 29 |
+
23 18.0961 ± 0.3583 -0.00482 ± 0.00229 0.03743 ± 0.00062 4.667 ± 0.094 % 90.501 ± 0.191 %
|
| 30 |
+
24 18.2480 ± 0.3546 -0.00422 ± 0.00226 0.03768 ± 0.00060 4.705 ± 0.091 % 90.376 ± 0.188 %
|
| 31 |
+
25 17.5764 ± 0.3333 -0.00442 ± 0.00218 0.03654 ± 0.00058 4.641 ± 0.089 % 90.588 ± 0.183 %
|
| 32 |
+
26 17.8519 ± 0.3317 -0.00528 ± 0.00213 0.03635 ± 0.00056 4.628 ± 0.087 % 90.567 ± 0.179 %
|
| 33 |
+
27 17.0149 ± 0.3091 -0.00536 ± 0.00210 0.03642 ± 0.00056 4.725 ± 0.089 % 90.681 ± 0.175 %
|
| 34 |
+
28 17.2258 ± 0.3069 -0.00531 ± 0.00205 0.03622 ± 0.00054 4.720 ± 0.086 % 90.661 ± 0.172 %
|
| 35 |
+
29 17.5403 ± 0.3081 -0.00458 ± 0.00203 0.03674 ± 0.00053 4.736 ± 0.084 % 90.559 ± 0.170 %
|
| 36 |
+
30 17.9357 ± 0.3095 -0.00474 ± 0.00201 0.03732 ± 0.00052 4.744 ± 0.081 % 90.443 ± 0.168 %
|
| 37 |
+
31 17.7863 ± 0.3006 -0.00374 ± 0.00197 0.03708 ± 0.00051 4.759 ± 0.081 % 90.483 ± 0.165 %
|
| 38 |
+
32 17.5273 ± 0.2914 -0.00405 ± 0.00193 0.03644 ± 0.00050 4.739 ± 0.079 % 90.619 ± 0.161 %
|
| 39 |
+
33 17.7630 ± 0.2905 -0.00468 ± 0.00190 0.03652 ± 0.00049 4.728 ± 0.077 % 90.521 ± 0.159 %
|
| 40 |
+
34 17.8153 ± 0.2863 -0.00406 ± 0.00186 0.03645 ± 0.00047 4.726 ± 0.076 % 90.417 ± 0.158 %
|
| 41 |
+
35 18.0621 ± 0.2868 -0.00338 ± 0.00185 0.03706 �� 0.00048 4.761 ± 0.075 % 90.353 ± 0.156 %
|
| 42 |
+
36 18.5182 ± 0.2913 -0.00470 ± 0.00264 0.07751 ± 0.00197 6.147 ± 0.122 % 89.109 ± 0.162 %
|
| 43 |
+
37 18.7423 ± 0.2907 -0.00474 ± 0.00258 0.07612 ± 0.00192 6.090 ± 0.120 % 89.205 ± 0.160 %
|
| 44 |
+
38 19.1342 ± 0.2933 -0.00552 ± 0.00253 0.07527 ± 0.00187 6.057 ± 0.118 % 89.126 ± 0.158 %
|
| 45 |
+
39 19.3702 ± 0.2930 -0.00508 ± 0.00248 0.07438 ± 0.00182 6.019 ± 0.116 % 89.109 ± 0.156 %
|
| 46 |
+
40 19.4712 ± 0.2904 -0.00574 ± 0.00243 0.07353 ± 0.00178 5.983 ± 0.114 % 89.074 ± 0.154 %
|
| 47 |
+
41 19.7055 ± 0.2906 -0.00579 ± 0.00239 0.07265 ± 0.00173 5.945 ± 0.112 % 89.071 ± 0.152 %
|
| 48 |
+
42 19.9599 ± 0.2912 -0.00584 ± 0.00235 0.07225 ± 0.00169 5.933 ± 0.109 % 89.010 ± 0.151 %
|
| 49 |
+
43 20.2370 ± 0.2918 -0.00689 ± 0.00231 0.07163 ± 0.00166 5.902 ± 0.108 % 88.997 ± 0.149 %
|
| 50 |
+
44 19.8878 ± 0.2826 -0.00680 ± 0.00226 0.07013 ± 0.00162 5.840 ± 0.106 % 89.161 ± 0.147 %
|
| 51 |
+
|
| 52 |
+
====== Perplexity statistics ======
|
| 53 |
+
Mean PPL(Q) : 19.887821 ± 0.282623
|
| 54 |
+
Mean PPL(base) : 20.023532 ± 0.285215
|
| 55 |
+
Cor(ln(PPL(Q)), ln(PPL(base))): 98.74%
|
| 56 |
+
Mean ln(PPL(Q)/PPL(base)) : -0.006801 ± 0.002259
|
| 57 |
+
Mean PPL(Q)/PPL(base) : 0.993222 ± 0.002243
|
| 58 |
+
Mean PPL(Q)-PPL(base) : -0.135711 ± 0.045139
|
| 59 |
+
|
| 60 |
+
====== KL divergence statistics ======
|
| 61 |
+
Mean KLD: 0.070127 ± 0.001619
|
| 62 |
+
Maximum KLD: 13.524951
|
| 63 |
+
99.9% KLD: 5.157535
|
| 64 |
+
99.0% KLD: 1.238282
|
| 65 |
+
95.0% KLD: 0.171232
|
| 66 |
+
90.0% KLD: 0.097467
|
| 67 |
+
Median KLD: 0.017696
|
| 68 |
+
10.0% KLD: 0.000818
|
| 69 |
+
5.0% KLD: 0.000128
|
| 70 |
+
1.0% KLD: 0.000003
|
| 71 |
+
0.1% KLD: -0.000002
|
| 72 |
+
Minimum KLD: -0.000086
|
| 73 |
+
|
| 74 |
+
====== Token probability statistics ======
|
| 75 |
+
Mean Δp: -0.018 ± 0.028 %
|
| 76 |
+
Maximum Δp: 99.446%
|
| 77 |
+
99.9% Δp: 53.868%
|
| 78 |
+
99.0% Δp: 16.426%
|
| 79 |
+
95.0% Δp: 6.337%
|
| 80 |
+
90.0% Δp: 3.328%
|
| 81 |
+
75.0% Δp: 0.540%
|
| 82 |
+
Median Δp: 0.000%
|
| 83 |
+
25.0% Δp: -0.562%
|
| 84 |
+
10.0% Δp: -3.485%
|
| 85 |
+
5.0% Δp: -6.557%
|
| 86 |
+
1.0% Δp: -17.290%
|
| 87 |
+
0.1% Δp: -49.032%
|
| 88 |
+
Minimum Δp: -99.623%
|
| 89 |
+
RMS Δp : 5.840 ± 0.106 %
|
| 90 |
+
Same top p: 89.161 ± 0.147 %
|
| 91 |
+
|
reports/kld2048-Q8_0.txt
ADDED
|
@@ -0,0 +1,91 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
0.04.058.683 I
|
| 2 |
+
0.04.058.704 I system_info: n_threads = 12 (n_threads_batch = 12) / 16 | MTL : EMBED_LIBRARY = 1 | CPU : NEON = 1 | ARM_FMA = 1 | FP16_VA = 1 | MATMUL_INT8 = 1 | DOTPROD = 1 | SME = 1 | ACCELERATE = 1 | OPENMP = 1 | REPACK = 1 |
|
| 3 |
+
0.04.074.455 I kl_divergence: computing over 44 chunks, n_ctx=2048, batch_size=2048, n_seq=1
|
| 4 |
+
0.07.642.683 I kl_divergence: 3.57 seconds per pass - ETA 2.62 minutes
|
| 5 |
+
|
| 6 |
+
chunk PPL ln(PPL(Q)/PPL(base)) KL Divergence Δp RMS Same top p
|
| 7 |
+
1 39.3179 ± 3.9687 0.00304 ± 0.00447 0.00512 ± 0.00020 1.719 ± 0.116 % 95.992 ± 0.614 %
|
| 8 |
+
2 43.7425 ± 3.1358 0.01101 ± 0.00482 0.01071 ± 0.00056 2.402 ± 0.135 % 94.379 ± 0.509 %
|
| 9 |
+
3 31.1566 ± 1.8236 0.00764 ± 0.00333 0.00775 ± 0.00040 2.043 ± 0.108 % 95.797 ± 0.362 %
|
| 10 |
+
4 21.3343 ± 1.0652 0.00590 ± 0.00256 0.00627 ± 0.00030 1.885 ± 0.090 % 96.359 ± 0.293 %
|
| 11 |
+
5 16.8234 ± 0.7337 0.00512 ± 0.00210 0.00519 ± 0.00025 1.726 ± 0.079 % 96.950 ± 0.240 %
|
| 12 |
+
6 19.0176 ± 0.7650 0.00615 ± 0.00192 0.00556 ± 0.00022 1.787 ± 0.071 % 96.693 ± 0.228 %
|
| 13 |
+
7 21.5324 ± 0.8110 0.00686 ± 0.00177 0.00567 ± 0.00019 1.783 ± 0.064 % 96.411 ± 0.220 %
|
| 14 |
+
8 23.7777 ± 0.8379 0.00781 ± 0.00163 0.00568 ± 0.00017 1.738 ± 0.058 % 96.261 ± 0.210 %
|
| 15 |
+
9 20.8430 ± 0.6883 0.00638 ± 0.00148 0.00523 ± 0.00015 1.671 ± 0.054 % 96.470 ± 0.192 %
|
| 16 |
+
10 23.4044 ± 0.7421 0.00727 ± 0.00140 0.00541 ± 0.00016 1.658 ± 0.050 % 96.325 ± 0.186 %
|
| 17 |
+
11 20.9095 ± 0.6296 0.00710 ± 0.00132 0.00510 ± 0.00015 1.624 ± 0.047 % 96.561 ± 0.172 %
|
| 18 |
+
12 20.1019 ± 0.5738 0.00644 ± 0.00124 0.00504 ± 0.00014 1.643 ± 0.045 % 96.562 ± 0.164 %
|
| 19 |
+
13 19.9460 ± 0.5443 0.00600 ± 0.00118 0.00507 ± 0.00013 1.675 ± 0.042 % 96.534 ± 0.159 %
|
| 20 |
+
14 21.1268 ± 0.5585 0.00622 ± 0.00116 0.00532 ± 0.00013 1.688 ± 0.040 % 96.327 ± 0.157 %
|
| 21 |
+
15 20.3644 ± 0.5177 0.00623 ± 0.00111 0.00519 ± 0.00012 1.697 ± 0.038 % 96.357 ± 0.151 %
|
| 22 |
+
16 18.8566 ± 0.4618 0.00571 ± 0.00105 0.00494 ± 0.00011 1.667 ± 0.037 % 96.518 ± 0.143 %
|
| 23 |
+
17 18.7770 ± 0.4422 0.00596 ± 0.00101 0.00486 ± 0.00011 1.653 ± 0.035 % 96.527 ± 0.139 %
|
| 24 |
+
18 19.3249 ± 0.4413 0.00632 ± 0.00096 0.00475 ± 0.00010 1.634 ± 0.034 % 96.573 ± 0.134 %
|
| 25 |
+
19 19.4971 ± 0.4323 0.00714 ± 0.00093 0.00471 ± 0.00010 1.629 ± 0.032 % 96.584 ± 0.130 %
|
| 26 |
+
20 19.9216 ± 0.4295 0.00719 ± 0.00090 0.00470 ± 0.00009 1.623 ± 0.031 % 96.608 ± 0.127 %
|
| 27 |
+
21 20.0633 ± 0.4215 0.00672 ± 0.00087 0.00462 ± 0.00009 1.614 ± 0.031 % 96.639 ± 0.123 %
|
| 28 |
+
22 19.3105 ± 0.3940 0.00646 ± 0.00084 0.00448 ± 0.00008 1.589 ± 0.030 % 96.716 ± 0.119 %
|
| 29 |
+
23 18.2955 ± 0.3633 0.00614 ± 0.00080 0.00433 ± 0.00008 1.565 ± 0.029 % 96.791 ± 0.115 %
|
| 30 |
+
24 18.4489 ± 0.3596 0.00673 ± 0.00081 0.00434 ± 0.00008 1.577 ± 0.028 % 96.790 ± 0.112 %
|
| 31 |
+
25 17.7665 ± 0.3379 0.00634 ± 0.00078 0.00421 ± 0.00007 1.556 ± 0.028 % 96.876 ± 0.109 %
|
| 32 |
+
26 18.0634 ± 0.3369 0.00650 ± 0.00076 0.00423 ± 0.00007 1.570 ± 0.027 % 96.857 ± 0.107 %
|
| 33 |
+
27 17.2203 ± 0.3140 0.00664 ± 0.00076 0.00431 ± 0.00008 1.617 ± 0.031 % 96.854 ± 0.105 %
|
| 34 |
+
28 17.4329 ± 0.3118 0.00664 ± 0.00074 0.00427 ± 0.00007 1.609 ± 0.030 % 96.837 ± 0.103 %
|
| 35 |
+
29 17.7293 ± 0.3125 0.00613 ± 0.00074 0.00435 ± 0.00007 1.622 ± 0.029 % 96.794 ± 0.102 %
|
| 36 |
+
30 18.1300 ± 0.3139 0.00604 ± 0.00073 0.00440 ± 0.00007 1.617 ± 0.028 % 96.774 ± 0.101 %
|
| 37 |
+
31 17.9560 ± 0.3043 0.00576 ± 0.00071 0.00436 ± 0.00007 1.617 ± 0.028 % 96.806 ± 0.099 %
|
| 38 |
+
32 17.7479 ± 0.2960 0.00845 ± 0.00077 0.00503 ± 0.00009 1.865 ± 0.047 % 96.722 ± 0.098 %
|
| 39 |
+
33 18.0002 ± 0.2953 0.00858 ± 0.00075 0.00500 ± 0.00009 1.857 ± 0.046 % 96.673 ± 0.098 %
|
| 40 |
+
34 18.0432 ± 0.2910 0.00865 ± 0.00074 0.00496 ± 0.00008 1.848 ± 0.045 % 96.651 ± 0.096 %
|
| 41 |
+
35 18.2803 ± 0.2911 0.00863 ± 0.00073 0.00500 �� 0.00008 1.848 ± 0.044 % 96.612 ± 0.096 %
|
| 42 |
+
36 18.7684 ± 0.2961 0.00872 ± 0.00079 0.00611 ± 0.00011 1.952 ± 0.044 % 96.416 ± 0.097 %
|
| 43 |
+
37 18.9962 ± 0.2956 0.00872 ± 0.00078 0.00603 ± 0.00011 1.934 ± 0.043 % 96.418 ± 0.096 %
|
| 44 |
+
38 19.4098 ± 0.2985 0.00878 ± 0.00076 0.00599 ± 0.00011 1.922 ± 0.042 % 96.393 ± 0.095 %
|
| 45 |
+
39 19.6427 ± 0.2980 0.00889 ± 0.00075 0.00593 ± 0.00011 1.911 ± 0.042 % 96.393 ± 0.093 %
|
| 46 |
+
40 19.7654 ± 0.2958 0.00925 ± 0.00073 0.00590 ± 0.00010 1.904 ± 0.041 % 96.378 ± 0.092 %
|
| 47 |
+
41 20.0024 ± 0.2960 0.00917 ± 0.00072 0.00587 ± 0.00010 1.894 ± 0.040 % 96.366 ± 0.091 %
|
| 48 |
+
42 20.2606 ± 0.2967 0.00911 ± 0.00071 0.00585 ± 0.00010 1.887 ± 0.039 % 96.348 ± 0.090 %
|
| 49 |
+
43 20.5612 ± 0.2977 0.00900 ± 0.00070 0.00583 ± 0.00010 1.879 ± 0.038 % 96.329 ± 0.090 %
|
| 50 |
+
44 20.2006 ± 0.2882 0.00880 ± 0.00069 0.00571 ± 0.00009 1.860 ± 0.038 % 96.381 ± 0.088 %
|
| 51 |
+
|
| 52 |
+
====== Perplexity statistics ======
|
| 53 |
+
Mean PPL(Q) : 20.200587 ± 0.288209
|
| 54 |
+
Mean PPL(base) : 20.023532 ± 0.285215
|
| 55 |
+
Cor(ln(PPL(Q)), ln(PPL(base))): 99.88%
|
| 56 |
+
Mean ln(PPL(Q)/PPL(base)) : 0.008803 ± 0.000686
|
| 57 |
+
Mean PPL(Q)/PPL(base) : 1.008842 ± 0.000692
|
| 58 |
+
Mean PPL(Q)-PPL(base) : 0.177055 ± 0.014106
|
| 59 |
+
|
| 60 |
+
====== KL divergence statistics ======
|
| 61 |
+
Mean KLD: 0.005708 ± 0.000095
|
| 62 |
+
Maximum KLD: 1.475488
|
| 63 |
+
99.9% KLD: 0.233686
|
| 64 |
+
99.0% KLD: 0.060422
|
| 65 |
+
95.0% KLD: 0.019779
|
| 66 |
+
90.0% KLD: 0.011459
|
| 67 |
+
Median KLD: 0.002024
|
| 68 |
+
10.0% KLD: 0.000093
|
| 69 |
+
5.0% KLD: 0.000014
|
| 70 |
+
1.0% KLD: -0.000000
|
| 71 |
+
0.1% KLD: -0.000027
|
| 72 |
+
Minimum KLD: -0.000139
|
| 73 |
+
|
| 74 |
+
====== Token probability statistics ======
|
| 75 |
+
Mean Δp: -0.129 ± 0.009 %
|
| 76 |
+
Maximum Δp: 26.548%
|
| 77 |
+
99.9% Δp: 12.262%
|
| 78 |
+
99.0% Δp: 4.915%
|
| 79 |
+
95.0% Δp: 1.852%
|
| 80 |
+
90.0% Δp: 0.953%
|
| 81 |
+
75.0% Δp: 0.123%
|
| 82 |
+
Median Δp: -0.001%
|
| 83 |
+
25.0% Δp: -0.252%
|
| 84 |
+
10.0% Δp: -1.292%
|
| 85 |
+
5.0% Δp: -2.469%
|
| 86 |
+
1.0% Δp: -6.287%
|
| 87 |
+
0.1% Δp: -15.758%
|
| 88 |
+
Minimum Δp: -56.803%
|
| 89 |
+
RMS Δp : 1.860 ± 0.038 %
|
| 90 |
+
Same top p: 96.381 ± 0.088 %
|
| 91 |
+
|
reports/kld2048b-IQ2_M.txt
ADDED
|
@@ -0,0 +1,91 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
0.01.613.406 I
|
| 2 |
+
0.01.613.424 I system_info: n_threads = 12 (n_threads_batch = 12) / 16 | MTL : EMBED_LIBRARY = 1 | CPU : NEON = 1 | ARM_FMA = 1 | FP16_VA = 1 | MATMUL_INT8 = 1 | DOTPROD = 1 | SME = 1 | ACCELERATE = 1 | OPENMP = 1 | REPACK = 1 |
|
| 3 |
+
0.01.629.925 I kl_divergence: computing over 44 chunks, n_ctx=2048, batch_size=2048, n_seq=1
|
| 4 |
+
0.05.557.685 I kl_divergence: 3.93 seconds per pass - ETA 2.87 minutes
|
| 5 |
+
|
| 6 |
+
chunk PPL ln(PPL(Q)/PPL(base)) KL Divergence Δp RMS Same top p
|
| 7 |
+
1 58.2738 ± 5.8636 0.39652 ± 0.04440 0.73148 ± 0.02548 17.754 ± 0.809 % 53.568 ± 1.560 %
|
| 8 |
+
2 28.1736 ± 1.9403 -0.42893 ± 0.04739 1.62751 ± 0.05135 27.095 ± 0.742 % 48.485 ± 1.105 %
|
| 9 |
+
3 24.1057 ± 1.3464 -0.24894 ± 0.03412 1.23011 ± 0.03642 23.821 ± 0.597 % 55.979 ± 0.896 %
|
| 10 |
+
4 18.1613 ± 0.8667 -0.15512 ± 0.02725 1.03879 ± 0.02884 22.542 ± 0.513 % 60.606 ± 0.764 %
|
| 11 |
+
5 15.2523 ± 0.6348 -0.09293 ± 0.02296 0.90438 ± 0.02390 21.904 ± 0.458 % 63.930 ± 0.672 %
|
| 12 |
+
6 18.4109 ± 0.7142 -0.02628 ± 0.02074 0.88345 ± 0.02052 21.337 ± 0.409 % 62.887 ± 0.617 %
|
| 13 |
+
7 22.3974 ± 0.8255 0.04625 ± 0.01934 0.86748 ± 0.01795 21.006 ± 0.371 % 61.346 ± 0.575 %
|
| 14 |
+
8 20.6588 ± 0.7044 -0.13280 ± 0.02011 1.13657 ± 0.02250 23.875 ± 0.373 % 58.798 ± 0.544 %
|
| 15 |
+
9 18.6723 ± 0.5957 -0.10359 ± 0.01844 1.06388 ± 0.02038 23.551 ± 0.352 % 60.302 ± 0.510 %
|
| 16 |
+
10 21.9508 ± 0.6830 -0.05685 ± 0.01731 1.03128 ± 0.01850 22.928 ± 0.331 % 59.638 ± 0.485 %
|
| 17 |
+
11 20.0201 ± 0.5899 -0.03637 ± 0.01640 0.99277 ± 0.01720 22.831 ± 0.314 % 60.739 ± 0.460 %
|
| 18 |
+
12 18.6809 ± 0.5225 -0.06687 ± 0.01563 1.01016 ± 0.01638 23.237 ± 0.298 % 60.834 ± 0.441 %
|
| 19 |
+
13 19.0078 ± 0.5098 -0.04218 ± 0.01478 0.97615 ± 0.01523 22.893 ± 0.283 % 60.944 ± 0.423 %
|
| 20 |
+
14 20.4682 ± 0.5313 -0.02545 ± 0.01420 0.96361 ± 0.01434 22.565 ± 0.271 % 60.529 ± 0.408 %
|
| 21 |
+
15 18.4287 ± 0.4568 -0.09364 ± 0.01414 1.03946 ± 0.01485 23.800 ± 0.265 % 60.169 ± 0.395 %
|
| 22 |
+
16 17.4307 ± 0.4156 -0.07292 ± 0.01355 1.00682 ± 0.01413 23.706 ± 0.257 % 60.997 ± 0.381 %
|
| 23 |
+
17 17.6917 ± 0.4072 -0.05359 ± 0.01300 0.98756 ± 0.01338 23.431 ± 0.248 % 60.882 ± 0.370 %
|
| 24 |
+
18 16.8473 ± 0.3737 -0.13088 ± 0.01310 1.06775 ± 0.01353 24.646 ± 0.244 % 60.041 ± 0.361 %
|
| 25 |
+
19 17.4819 ± 0.3780 -0.10195 ± 0.01263 1.04620 ± 0.01290 24.367 ± 0.236 % 59.958 ± 0.351 %
|
| 26 |
+
20 18.3668 ± 0.3873 -0.07407 ± 0.01226 1.03257 ± 0.01232 24.077 ± 0.229 % 59.780 ± 0.343 %
|
| 27 |
+
21 18.8620 ± 0.3886 -0.05502 ± 0.01190 1.01193 ± 0.01180 23.846 ± 0.222 % 59.791 ± 0.335 %
|
| 28 |
+
22 18.3624 ± 0.3674 -0.04388 ± 0.01146 0.98364 ± 0.01133 23.573 ± 0.216 % 60.375 ± 0.326 %
|
| 29 |
+
23 17.5539 ± 0.3417 -0.03524 ± 0.01107 0.95691 ± 0.01090 23.362 ± 0.211 % 61.001 ± 0.318 %
|
| 30 |
+
24 18.0106 ± 0.3447 -0.01731 ± 0.01082 0.94846 ± 0.01053 23.198 ± 0.206 % 60.952 ± 0.311 %
|
| 31 |
+
25 17.4314 ± 0.3253 -0.01270 ± 0.01045 0.92335 ± 0.01015 22.984 ± 0.201 % 61.529 ± 0.304 %
|
| 32 |
+
26 17.9211 ± 0.3279 -0.00141 ± 0.01018 0.90906 ± 0.00979 22.789 ± 0.196 % 61.456 ± 0.298 %
|
| 33 |
+
27 17.2064 ± 0.3074 0.00583 ± 0.00992 0.89507 ± 0.00951 22.838 ± 0.192 % 61.859 ± 0.292 %
|
| 34 |
+
28 17.6901 ± 0.3111 0.02129 ± 0.00969 0.88479 ± 0.00921 22.665 ± 0.188 % 61.877 ± 0.287 %
|
| 35 |
+
29 18.2022 ± 0.3160 0.03246 ± 0.00952 0.88060 ± 0.00895 22.602 ± 0.184 % 61.691 ± 0.282 %
|
| 36 |
+
30 18.8616 ± 0.3227 0.04560 ± 0.00936 0.87818 ± 0.00871 22.473 ± 0.180 % 61.492 ± 0.278 %
|
| 37 |
+
31 18.7717 ± 0.3146 0.05019 ± 0.00914 0.86753 ± 0.00846 22.370 ± 0.177 % 61.590 ± 0.273 %
|
| 38 |
+
32 18.8487 ± 0.3107 0.06863 ± 0.00907 0.87097 ± 0.00828 22.641 ± 0.175 % 61.431 ± 0.269 %
|
| 39 |
+
33 19.2552 ± 0.3123 0.07598 ± 0.00889 0.86278 ± 0.00806 22.499 ± 0.172 % 61.358 ± 0.265 %
|
| 40 |
+
34 19.5123 ± 0.3115 0.08693 ± 0.00871 0.85368 ± 0.00785 22.374 ± 0.168 % 61.287 ± 0.261 %
|
| 41 |
+
35 19.9289 ± 0.3141 0.09497 ± 0.00857 0.84976 �� 0.00765 22.335 ± 0.165 % 61.170 ± 0.258 %
|
| 42 |
+
36 19.7929 ± 0.3065 0.06187 ± 0.00864 0.88110 ± 0.00775 22.485 ± 0.163 % 60.549 ± 0.255 %
|
| 43 |
+
37 20.0909 ± 0.3068 0.06474 ± 0.00846 0.87049 ± 0.00756 22.319 ± 0.160 % 60.617 ± 0.251 %
|
| 44 |
+
38 20.6278 ± 0.3115 0.06964 ± 0.00831 0.86380 ± 0.00739 22.192 ± 0.158 % 60.441 ± 0.248 %
|
| 45 |
+
39 21.0228 ± 0.3135 0.07679 ± 0.00817 0.85589 ± 0.00721 22.061 ± 0.155 % 60.431 ± 0.245 %
|
| 46 |
+
40 21.2105 ± 0.3123 0.07982 ± 0.00803 0.84951 ± 0.00706 21.953 ± 0.153 % 60.411 ± 0.242 %
|
| 47 |
+
41 21.5474 ± 0.3138 0.08357 ± 0.00790 0.84404 ± 0.00691 21.850 ± 0.151 % 60.401 ± 0.239 %
|
| 48 |
+
42 21.9250 ± 0.3162 0.08806 ± 0.00779 0.83956 ± 0.00677 21.753 ± 0.148 % 60.364 ± 0.236 %
|
| 49 |
+
43 22.3737 ± 0.3191 0.09349 ± 0.00767 0.83571 ± 0.00664 21.655 ± 0.146 % 60.299 ± 0.233 %
|
| 50 |
+
44 21.9919 ± 0.3091 0.09377 ± 0.00751 0.82111 ± 0.00651 21.470 ± 0.145 % 60.690 ± 0.230 %
|
| 51 |
+
|
| 52 |
+
====== Perplexity statistics ======
|
| 53 |
+
Mean PPL(Q) : 21.991922 ± 0.309087
|
| 54 |
+
Mean PPL(base) : 20.023532 ± 0.285215
|
| 55 |
+
Cor(ln(PPL(Q)), ln(PPL(base))): 85.91%
|
| 56 |
+
Mean ln(PPL(Q)/PPL(base)) : 0.093767 ± 0.007514
|
| 57 |
+
Mean PPL(Q)/PPL(base) : 1.098304 ± 0.008253
|
| 58 |
+
Mean PPL(Q)-PPL(base) : 1.968390 ± 0.159436
|
| 59 |
+
|
| 60 |
+
====== KL divergence statistics ======
|
| 61 |
+
Mean KLD: 0.821109 ± 0.006506
|
| 62 |
+
Maximum KLD: 27.317856
|
| 63 |
+
99.9% KLD: 13.975519
|
| 64 |
+
99.0% KLD: 7.225517
|
| 65 |
+
95.0% KLD: 2.905154
|
| 66 |
+
90.0% KLD: 1.848341
|
| 67 |
+
Median KLD: 0.419832
|
| 68 |
+
10.0% KLD: 0.037137
|
| 69 |
+
5.0% KLD: 0.008512
|
| 70 |
+
1.0% KLD: 0.000367
|
| 71 |
+
0.1% KLD: 0.000015
|
| 72 |
+
Minimum KLD: -0.000003
|
| 73 |
+
|
| 74 |
+
====== Token probability statistics ======
|
| 75 |
+
Mean Δp: -2.527 ± 0.100 %
|
| 76 |
+
Maximum Δp: 99.984%
|
| 77 |
+
99.9% Δp: 99.142%
|
| 78 |
+
99.0% Δp: 77.508%
|
| 79 |
+
95.0% Δp: 27.281%
|
| 80 |
+
90.0% Δp: 12.375%
|
| 81 |
+
75.0% Δp: 1.135%
|
| 82 |
+
Median Δp: -0.116%
|
| 83 |
+
25.0% Δp: -6.481%
|
| 84 |
+
10.0% Δp: -23.965%
|
| 85 |
+
5.0% Δp: -38.637%
|
| 86 |
+
1.0% Δp: -72.888%
|
| 87 |
+
0.1% Δp: -96.885%
|
| 88 |
+
Minimum Δp: -99.975%
|
| 89 |
+
RMS Δp : 21.470 ± 0.145 %
|
| 90 |
+
Same top p: 60.690 ± 0.230 %
|
| 91 |
+
|
reports/kld2048b-IQ4_XS.txt
ADDED
|
@@ -0,0 +1,91 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
0.02.378.266 I
|
| 2 |
+
0.02.378.288 I system_info: n_threads = 12 (n_threads_batch = 12) / 16 | MTL : EMBED_LIBRARY = 1 | CPU : NEON = 1 | ARM_FMA = 1 | FP16_VA = 1 | MATMUL_INT8 = 1 | DOTPROD = 1 | SME = 1 | ACCELERATE = 1 | OPENMP = 1 | REPACK = 1 |
|
| 3 |
+
0.02.392.915 I kl_divergence: computing over 44 chunks, n_ctx=2048, batch_size=2048, n_seq=1
|
| 4 |
+
0.05.866.452 I kl_divergence: 3.47 seconds per pass - ETA 2.53 minutes
|
| 5 |
+
|
| 6 |
+
chunk PPL ln(PPL(Q)/PPL(base)) KL Divergence Δp RMS Same top p
|
| 7 |
+
1 40.8014 ± 4.0883 0.04008 ± 0.02030 0.13138 ± 0.00527 7.765 ± 0.478 % 76.637 ± 1.324 %
|
| 8 |
+
2 46.1092 ± 3.3048 0.06370 ± 0.01516 0.13641 ± 0.00396 8.106 ± 0.383 % 77.713 ± 0.920 %
|
| 9 |
+
3 32.8071 ± 1.9299 0.05926 ± 0.01117 0.11340 ± 0.00316 7.545 ± 0.350 % 81.069 ± 0.707 %
|
| 10 |
+
4 22.4515 ± 1.1303 0.05694 ± 0.00948 0.10398 ± 0.00281 7.428 ± 0.296 % 83.431 ± 0.581 %
|
| 11 |
+
5 17.6272 ± 0.7763 0.05179 ± 0.00817 0.09287 ± 0.00269 7.189 ± 0.283 % 85.415 ± 0.494 %
|
| 12 |
+
6 19.9152 ± 0.8074 0.05226 ± 0.00775 0.10057 ± 0.00249 7.343 ± 0.245 % 84.572 ± 0.461 %
|
| 13 |
+
7 22.5669 ± 0.8587 0.05378 ± 0.00731 0.10573 ± 0.00230 7.330 ± 0.220 % 83.675 ± 0.437 %
|
| 14 |
+
8 24.8553 ± 0.8845 0.05213 ± 0.00703 0.11585 ± 0.00230 7.539 ± 0.207 % 82.564 ± 0.419 %
|
| 15 |
+
9 21.6370 ± 0.7205 0.04377 ± 0.00651 0.11104 ± 0.00220 7.532 ± 0.195 % 83.263 ± 0.389 %
|
| 16 |
+
10 24.0920 ± 0.7686 0.03623 ± 0.00625 0.11825 ± 0.00211 7.547 ± 0.183 % 82.444 ± 0.376 %
|
| 17 |
+
11 21.4854 ± 0.6509 0.03427 ± 0.00593 0.11635 ± 0.00203 7.631 ± 0.176 % 82.982 ± 0.354 %
|
| 18 |
+
12 20.5664 ± 0.5904 0.02929 ± 0.00562 0.11540 ± 0.00191 7.732 ± 0.165 % 83.170 ± 0.338 %
|
| 19 |
+
13 20.2696 ± 0.5561 0.02209 ± 0.00541 0.11621 ± 0.00181 7.861 ± 0.156 % 82.886 ± 0.327 %
|
| 20 |
+
14 21.4866 ± 0.5707 0.02311 ± 0.00524 0.11889 ± 0.00175 7.874 ± 0.150 % 82.642 ± 0.316 %
|
| 21 |
+
15 19.1890 ± 0.4870 -0.05322 ± 0.00702 0.25911 ± 0.00767 12.351 ± 0.258 % 80.828 ± 0.318 %
|
| 22 |
+
16 17.8779 ± 0.4370 -0.04759 ± 0.00666 0.24714 ± 0.00722 12.113 ± 0.249 % 81.433 ± 0.304 %
|
| 23 |
+
17 17.9925 ± 0.4230 -0.03672 ± 0.00639 0.24029 ± 0.00681 11.982 ± 0.238 % 81.347 ± 0.295 %
|
| 24 |
+
18 18.6280 ± 0.4250 -0.03041 ± 0.00610 0.23248 ± 0.00644 11.771 ± 0.229 % 81.340 ± 0.287 %
|
| 25 |
+
19 18.9205 ± 0.4190 -0.02287 ± 0.00588 0.22702 ± 0.00610 11.628 ± 0.220 % 81.309 ± 0.280 %
|
| 26 |
+
20 19.4109 ± 0.4180 -0.01878 ± 0.00568 0.22270 ± 0.00581 11.457 ± 0.213 % 81.251 ± 0.273 %
|
| 27 |
+
21 19.6166 ± 0.4115 -0.01580 ± 0.00546 0.21691 ± 0.00554 11.292 ± 0.206 % 81.409 ± 0.265 %
|
| 28 |
+
22 18.9276 ± 0.3856 -0.01356 ± 0.00525 0.20974 ± 0.00529 11.132 ± 0.200 % 81.778 ± 0.257 %
|
| 29 |
+
23 17.9744 ± 0.3565 -0.01156 ± 0.00506 0.20318 ± 0.00507 11.028 ± 0.195 % 82.150 ± 0.250 %
|
| 30 |
+
24 18.2363 ± 0.3555 -0.00486 ± 0.00496 0.20124 ± 0.00487 11.007 ± 0.189 % 82.030 ± 0.245 %
|
| 31 |
+
25 17.5697 ± 0.3343 -0.00480 ± 0.00478 0.19488 ± 0.00468 10.837 ± 0.185 % 82.389 ± 0.238 %
|
| 32 |
+
26 17.9133 ± 0.3342 -0.00184 ± 0.00465 0.19186 ± 0.00451 10.751 ± 0.180 % 82.356 ± 0.234 %
|
| 33 |
+
27 17.1145 ± 0.3124 0.00048 ± 0.00457 0.19081 ± 0.00438 10.812 ± 0.175 % 82.470 ± 0.229 %
|
| 34 |
+
28 17.3774 ± 0.3113 0.00346 ± 0.00445 0.18779 ± 0.00423 10.712 ± 0.171 % 82.495 ± 0.225 %
|
| 35 |
+
29 17.7102 ± 0.3128 0.00505 ± 0.00437 0.18741 ± 0.00410 10.681 ± 0.166 % 82.354 ± 0.221 %
|
| 36 |
+
30 18.1273 ± 0.3146 0.00589 ± 0.00429 0.18601 ± 0.00396 10.586 ± 0.163 % 82.255 ± 0.218 %
|
| 37 |
+
31 18.0144 ± 0.3063 0.00901 ± 0.00418 0.18324 ± 0.00384 10.512 ± 0.159 % 82.304 ± 0.214 %
|
| 38 |
+
32 18.3101 ± 0.3064 0.03964 ± 0.00475 0.22681 ± 0.00426 12.223 ± 0.173 % 81.128 ± 0.216 %
|
| 39 |
+
33 18.5720 ± 0.3056 0.03985 ± 0.00464 0.22342 ± 0.00414 12.120 ± 0.170 % 81.078 ± 0.213 %
|
| 40 |
+
34 18.6240 ± 0.3012 0.04034 ± 0.00454 0.22017 ± 0.00402 12.009 ± 0.166 % 81.059 ± 0.210 %
|
| 41 |
+
35 18.9015 ± 0.3020 0.04204 ± 0.00445 0.21834 �� 0.00391 11.930 ± 0.163 % 81.045 ± 0.207 %
|
| 42 |
+
36 19.3547 ± 0.3059 0.03948 ± 0.00464 0.24281 ± 0.00405 12.260 ± 0.160 % 80.184 ± 0.208 %
|
| 43 |
+
37 19.5866 ± 0.3052 0.03932 ± 0.00453 0.23854 ± 0.00395 12.143 ± 0.158 % 80.312 ± 0.204 %
|
| 44 |
+
38 20.0163 ± 0.3081 0.03955 ± 0.00443 0.23544 ± 0.00385 12.042 ± 0.155 % 80.257 ± 0.202 %
|
| 45 |
+
39 20.2668 ± 0.3077 0.04017 ± 0.00434 0.23233 ± 0.00375 11.937 ± 0.152 % 80.264 ± 0.199 %
|
| 46 |
+
40 20.3846 ± 0.3054 0.04010 ± 0.00425 0.22910 ± 0.00366 11.841 ± 0.150 % 80.347 ± 0.196 %
|
| 47 |
+
41 20.6113 ± 0.3052 0.03915 ± 0.00417 0.22629 ± 0.00357 11.752 ± 0.148 % 80.335 ± 0.194 %
|
| 48 |
+
42 20.8683 ± 0.3056 0.03866 ± 0.00410 0.22436 ± 0.00349 11.685 ± 0.145 % 80.296 ± 0.192 %
|
| 49 |
+
43 21.1594 ± 0.3063 0.03768 ± 0.00403 0.22184 ± 0.00341 11.595 ± 0.143 % 80.309 ± 0.190 %
|
| 50 |
+
44 20.7806 ± 0.2965 0.03711 ± 0.00395 0.21727 ± 0.00334 11.474 ± 0.141 % 80.554 ± 0.187 %
|
| 51 |
+
|
| 52 |
+
====== Perplexity statistics ======
|
| 53 |
+
Mean PPL(Q) : 20.780607 ± 0.296500
|
| 54 |
+
Mean PPL(base) : 20.023532 ± 0.285215
|
| 55 |
+
Cor(ln(PPL(Q)), ln(PPL(base))): 96.17%
|
| 56 |
+
Mean ln(PPL(Q)/PPL(base)) : 0.037112 ± 0.003946
|
| 57 |
+
Mean PPL(Q)/PPL(base) : 1.037809 ± 0.004096
|
| 58 |
+
Mean PPL(Q)-PPL(base) : 0.757076 ± 0.081286
|
| 59 |
+
|
| 60 |
+
====== KL divergence statistics ======
|
| 61 |
+
Mean KLD: 0.217275 ± 0.003340
|
| 62 |
+
Maximum KLD: 17.964737
|
| 63 |
+
99.9% KLD: 9.634006
|
| 64 |
+
99.0% KLD: 3.159044
|
| 65 |
+
95.0% KLD: 0.760046
|
| 66 |
+
90.0% KLD: 0.369870
|
| 67 |
+
Median KLD: 0.066936
|
| 68 |
+
10.0% KLD: 0.003958
|
| 69 |
+
5.0% KLD: 0.000743
|
| 70 |
+
1.0% KLD: 0.000018
|
| 71 |
+
0.1% KLD: 0.000000
|
| 72 |
+
Minimum KLD: -0.000052
|
| 73 |
+
|
| 74 |
+
====== Token probability statistics ======
|
| 75 |
+
Mean Δp: -0.564 ± 0.054 %
|
| 76 |
+
Maximum Δp: 99.982%
|
| 77 |
+
99.9% Δp: 91.608%
|
| 78 |
+
99.0% Δp: 34.479%
|
| 79 |
+
95.0% Δp: 11.713%
|
| 80 |
+
90.0% Δp: 5.915%
|
| 81 |
+
75.0% Δp: 0.778%
|
| 82 |
+
Median Δp: -0.004%
|
| 83 |
+
25.0% Δp: -1.506%
|
| 84 |
+
10.0% Δp: -7.762%
|
| 85 |
+
5.0% Δp: -14.633%
|
| 86 |
+
1.0% Δp: -39.807%
|
| 87 |
+
0.1% Δp: -93.063%
|
| 88 |
+
Minimum Δp: -99.936%
|
| 89 |
+
RMS Δp : 11.474 ± 0.141 %
|
| 90 |
+
Same top p: 80.554 ± 0.187 %
|
| 91 |
+
|
reports/kld2048b-Q3_K_M.txt
ADDED
|
@@ -0,0 +1,91 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
0.02.201.588 I
|
| 2 |
+
0.02.201.608 I system_info: n_threads = 12 (n_threads_batch = 12) / 16 | MTL : EMBED_LIBRARY = 1 | CPU : NEON = 1 | ARM_FMA = 1 | FP16_VA = 1 | MATMUL_INT8 = 1 | DOTPROD = 1 | SME = 1 | ACCELERATE = 1 | OPENMP = 1 | REPACK = 1 |
|
| 3 |
+
0.02.216.355 I kl_divergence: computing over 44 chunks, n_ctx=2048, batch_size=2048, n_seq=1
|
| 4 |
+
0.06.094.651 I kl_divergence: 3.88 seconds per pass - ETA 2.83 minutes
|
| 5 |
+
|
| 6 |
+
chunk PPL ln(PPL(Q)/PPL(base)) KL Divergence Δp RMS Same top p
|
| 7 |
+
1 44.0050 ± 4.3037 0.11567 ± 0.02992 0.28583 ± 0.01011 11.510 ± 0.624 % 68.817 ± 1.449 %
|
| 8 |
+
2 49.0206 ± 3.5001 0.12493 ± 0.02173 0.28682 ± 0.00748 11.258 ± 0.444 % 70.430 ± 1.009 %
|
| 9 |
+
3 34.3613 ± 2.0141 0.10554 ± 0.01647 0.24723 ± 0.00612 10.763 ± 0.392 % 74.356 ± 0.788 %
|
| 10 |
+
4 23.0288 ± 1.1546 0.08233 ± 0.01416 0.23331 ± 0.00578 11.082 ± 0.354 % 76.808 ± 0.660 %
|
| 11 |
+
5 18.0180 ± 0.7893 0.07372 ± 0.01208 0.20844 ± 0.00508 11.076 ± 0.339 % 79.179 ± 0.568 %
|
| 12 |
+
6 20.7834 ± 0.8412 0.09494 ± 0.01133 0.23622 ± 0.00501 11.508 ± 0.308 % 77.452 ± 0.533 %
|
| 13 |
+
7 23.7445 ± 0.8998 0.10465 ± 0.01061 0.24902 ± 0.00473 11.596 ± 0.281 % 76.316 ± 0.502 %
|
| 14 |
+
8 26.1189 ± 0.9268 0.10172 ± 0.01016 0.26971 ± 0.00457 11.818 ± 0.261 % 74.548 ± 0.482 %
|
| 15 |
+
9 22.7897 ± 0.7579 0.09567 ± 0.00959 0.26207 ± 0.00454 12.018 ± 0.261 % 75.627 ± 0.447 %
|
| 16 |
+
10 25.5304 ± 0.8149 0.09422 ± 0.00910 0.26722 ± 0.00423 11.875 ± 0.245 % 74.673 ± 0.430 %
|
| 17 |
+
11 22.6517 ± 0.6841 0.08713 ± 0.00868 0.26373 ± 0.00413 12.127 ± 0.239 % 75.473 ± 0.406 %
|
| 18 |
+
12 21.9205 ± 0.6291 0.09305 ± 0.00824 0.25784 ± 0.00386 12.107 ± 0.224 % 75.896 ± 0.386 %
|
| 19 |
+
13 21.8487 ± 0.6001 0.09711 ± 0.00794 0.26151 ± 0.00369 12.217 ± 0.212 % 75.389 ± 0.374 %
|
| 20 |
+
14 23.1626 ± 0.6156 0.09822 ± 0.00783 0.27225 ± 0.00376 12.233 ± 0.202 % 74.787 ± 0.363 %
|
| 21 |
+
15 22.3213 ± 0.5704 0.09799 ± 0.00755 0.26907 ± 0.00358 12.364 ± 0.194 % 74.852 ± 0.350 %
|
| 22 |
+
16 20.5861 ± 0.5066 0.09346 ± 0.00726 0.26292 ± 0.00350 12.466 ± 0.191 % 75.538 ± 0.336 %
|
| 23 |
+
17 20.4422 ± 0.4849 0.09092 ± 0.00705 0.26347 ± 0.00334 12.436 ± 0.182 % 75.418 ± 0.327 %
|
| 24 |
+
18 21.1229 ± 0.4864 0.09528 ± 0.00677 0.26083 ± 0.00318 12.345 ± 0.175 % 75.220 ± 0.318 %
|
| 25 |
+
19 21.4692 ± 0.4813 0.10350 ± 0.00664 0.26323 ± 0.00309 12.353 ± 0.169 % 75.073 ± 0.310 %
|
| 26 |
+
20 21.9715 ± 0.4797 0.10513 ± 0.00652 0.26749 ± 0.00300 12.335 ± 0.163 % 74.795 ± 0.304 %
|
| 27 |
+
21 22.3093 ± 0.4762 0.11283 ± 0.00637 0.26714 ± 0.00289 12.254 ± 0.158 % 74.831 ± 0.296 %
|
| 28 |
+
22 21.4290 ± 0.4442 0.11056 ± 0.00619 0.26275 ± 0.00288 12.236 ± 0.156 % 75.162 ± 0.288 %
|
| 29 |
+
23 20.2528 ± 0.4087 0.10778 ± 0.00599 0.25773 ± 0.00281 12.209 ± 0.153 % 75.520 ± 0.280 %
|
| 30 |
+
24 20.3877 ± 0.4037 0.10666 ± 0.00591 0.25995 ± 0.00278 12.223 ± 0.149 % 75.411 ± 0.275 %
|
| 31 |
+
25 19.6101 ± 0.3790 0.10507 ± 0.00572 0.25423 ± 0.00269 12.112 ± 0.145 % 75.844 ± 0.268 %
|
| 32 |
+
26 19.9641 ± 0.3780 0.10654 ± 0.00559 0.25261 ± 0.00260 12.072 ± 0.142 % 75.761 ± 0.263 %
|
| 33 |
+
27 18.9935 ± 0.3515 0.10465 ± 0.00548 0.25139 ± 0.00259 12.157 ± 0.140 % 76.025 ± 0.257 %
|
| 34 |
+
28 19.2026 ± 0.3484 0.10333 ± 0.00536 0.25101 ± 0.00253 12.132 ± 0.137 % 75.971 ± 0.252 %
|
| 35 |
+
29 19.5907 ± 0.3502 0.10597 ± 0.00529 0.25350 ± 0.00249 12.208 ± 0.135 % 75.798 ± 0.249 %
|
| 36 |
+
30 20.0606 ± 0.3525 0.10723 ± 0.00523 0.25618 ± 0.00246 12.194 ± 0.133 % 75.614 ± 0.245 %
|
| 37 |
+
31 19.8650 ± 0.3421 0.10679 ± 0.00514 0.25552 ± 0.00240 12.193 ± 0.130 % 75.600 ± 0.241 %
|
| 38 |
+
32 20.1615 ± 0.3426 0.13596 ± 0.00550 0.28938 ± 0.00299 13.417 ± 0.147 % 74.823 ± 0.240 %
|
| 39 |
+
33 20.4302 ± 0.3413 0.13522 ± 0.00541 0.28839 ± 0.00292 13.350 ± 0.144 % 74.819 ± 0.236 %
|
| 40 |
+
34 20.4814 ± 0.3362 0.13540 ± 0.00530 0.28669 ± 0.00284 13.315 ± 0.141 % 74.760 ± 0.233 %
|
| 41 |
+
35 20.7382 ± 0.3361 0.13478 ± 0.00524 0.28906 �� 0.00279 13.364 ± 0.138 % 74.571 ± 0.230 %
|
| 42 |
+
36 20.4010 ± 0.3249 0.09213 ± 0.00569 0.34290 ± 0.00388 14.213 ± 0.142 % 73.626 ± 0.230 %
|
| 43 |
+
37 20.6839 ± 0.3250 0.09383 ± 0.00557 0.33923 ± 0.00378 14.116 ± 0.140 % 73.626 ± 0.227 %
|
| 44 |
+
38 21.1550 ± 0.3287 0.09488 ± 0.00547 0.33794 ± 0.00370 14.049 ± 0.137 % 73.553 ± 0.224 %
|
| 45 |
+
39 21.4587 ± 0.3293 0.09731 ± 0.00537 0.33523 ± 0.00361 13.966 ± 0.135 % 73.587 ± 0.221 %
|
| 46 |
+
40 21.6332 ± 0.3279 0.09955 ± 0.00528 0.33395 ± 0.00353 13.916 ± 0.133 % 73.541 ± 0.218 %
|
| 47 |
+
41 21.9511 ± 0.3292 0.10213 ± 0.00521 0.33273 ± 0.00345 13.851 ± 0.131 % 73.421 ± 0.216 %
|
| 48 |
+
42 22.3018 ± 0.3313 0.10510 ± 0.00513 0.33257 ± 0.00338 13.804 ± 0.129 % 73.351 ± 0.213 %
|
| 49 |
+
43 22.6720 ± 0.3332 0.10673 ± 0.00507 0.33194 ± 0.00333 13.755 ± 0.127 % 73.307 ± 0.211 %
|
| 50 |
+
44 22.2382 ± 0.3220 0.10490 ± 0.00496 0.32583 ± 0.00326 13.630 ± 0.125 % 73.580 ± 0.208 %
|
| 51 |
+
|
| 52 |
+
====== Perplexity statistics ======
|
| 53 |
+
Mean PPL(Q) : 22.238185 ± 0.322012
|
| 54 |
+
Mean PPL(base) : 20.023532 ± 0.285215
|
| 55 |
+
Cor(ln(PPL(Q)), ln(PPL(base))): 94.05%
|
| 56 |
+
Mean ln(PPL(Q)/PPL(base)) : 0.104903 ± 0.004960
|
| 57 |
+
Mean PPL(Q)/PPL(base) : 1.110603 ± 0.005509
|
| 58 |
+
Mean PPL(Q)-PPL(base) : 2.214654 ± 0.110837
|
| 59 |
+
|
| 60 |
+
====== KL divergence statistics ======
|
| 61 |
+
Mean KLD: 0.325828 ± 0.003258
|
| 62 |
+
Maximum KLD: 19.564314
|
| 63 |
+
99.9% KLD: 8.606861
|
| 64 |
+
99.0% KLD: 3.195640
|
| 65 |
+
95.0% KLD: 1.148634
|
| 66 |
+
90.0% KLD: 0.695183
|
| 67 |
+
Median KLD: 0.149037
|
| 68 |
+
10.0% KLD: 0.009316
|
| 69 |
+
5.0% KLD: 0.001733
|
| 70 |
+
1.0% KLD: 0.000054
|
| 71 |
+
0.1% KLD: 0.000001
|
| 72 |
+
Minimum KLD: -0.000010
|
| 73 |
+
|
| 74 |
+
====== Token probability statistics ======
|
| 75 |
+
Mean Δp: -1.203 ± 0.064 %
|
| 76 |
+
Maximum Δp: 99.921%
|
| 77 |
+
99.9% Δp: 83.312%
|
| 78 |
+
99.0% Δp: 40.107%
|
| 79 |
+
95.0% Δp: 15.836%
|
| 80 |
+
90.0% Δp: 8.091%
|
| 81 |
+
75.0% Δp: 1.037%
|
| 82 |
+
Median Δp: -0.017%
|
| 83 |
+
25.0% Δp: -2.502%
|
| 84 |
+
10.0% Δp: -12.324%
|
| 85 |
+
5.0% Δp: -22.158%
|
| 86 |
+
1.0% Δp: -50.154%
|
| 87 |
+
0.1% Δp: -91.656%
|
| 88 |
+
Minimum Δp: -99.845%
|
| 89 |
+
RMS Δp : 13.630 ± 0.125 %
|
| 90 |
+
Same top p: 73.580 ± 0.208 %
|
| 91 |
+
|
reports/kld2048b-Q4_K_M.txt
ADDED
|
@@ -0,0 +1,91 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
0.02.079.698 I
|
| 2 |
+
0.02.079.719 I system_info: n_threads = 12 (n_threads_batch = 12) / 16 | MTL : EMBED_LIBRARY = 1 | CPU : NEON = 1 | ARM_FMA = 1 | FP16_VA = 1 | MATMUL_INT8 = 1 | DOTPROD = 1 | SME = 1 | ACCELERATE = 1 | OPENMP = 1 | REPACK = 1 |
|
| 3 |
+
0.02.093.411 I kl_divergence: computing over 44 chunks, n_ctx=2048, batch_size=2048, n_seq=1
|
| 4 |
+
0.06.162.087 I kl_divergence: 4.07 seconds per pass - ETA 2.98 minutes
|
| 5 |
+
|
| 6 |
+
chunk PPL ln(PPL(Q)/PPL(base)) KL Divergence Δp RMS Same top p
|
| 7 |
+
1 41.2702 ± 4.1811 0.05150 ± 0.02115 0.14859 ± 0.00665 8.046 ± 0.420 % 78.592 ± 1.283 %
|
| 8 |
+
2 44.5736 ± 3.1803 0.02983 ± 0.01537 0.14332 ± 0.00463 8.115 ± 0.383 % 78.983 ± 0.901 %
|
| 9 |
+
3 31.3973 ± 1.8245 0.01533 ± 0.01133 0.11695 ± 0.00374 7.745 ± 0.367 % 82.470 ± 0.686 %
|
| 10 |
+
4 21.5957 ± 1.0734 0.01808 ± 0.00942 0.10860 ± 0.00433 7.784 ± 0.335 % 84.384 ± 0.568 %
|
| 11 |
+
5 16.9509 ± 0.7364 0.01266 ± 0.00797 0.09431 ± 0.00355 7.447 ± 0.290 % 86.276 ± 0.481 %
|
| 12 |
+
6 19.4643 ± 0.7848 0.02936 ± 0.00782 0.10815 ± 0.00334 7.874 ± 0.266 % 84.848 ± 0.458 %
|
| 13 |
+
7 22.6163 ± 0.8587 0.05597 ± 0.00756 0.11632 ± 0.00305 7.995 ± 0.240 % 83.689 ± 0.437 %
|
| 14 |
+
8 20.5296 ± 0.7191 -0.13907 ± 0.01298 0.46931 ± 0.01663 15.957 ± 0.400 % 78.372 ± 0.455 %
|
| 15 |
+
9 18.3012 ± 0.5997 -0.12367 ± 0.01168 0.42440 ± 0.01486 15.241 ± 0.375 % 79.668 ± 0.419 %
|
| 16 |
+
10 20.9340 ± 0.6618 -0.10428 ± 0.01075 0.39735 ± 0.01342 14.637 ± 0.354 % 79.492 ± 0.399 %
|
| 17 |
+
11 18.9142 ± 0.5668 -0.09319 ± 0.00991 0.36823 ± 0.01224 14.119 ± 0.334 % 80.494 ± 0.374 %
|
| 18 |
+
12 19.0460 ± 0.5453 -0.04751 ± 0.00953 0.36436 ± 0.01130 14.121 ± 0.314 % 80.173 ± 0.360 %
|
| 19 |
+
13 18.9689 ± 0.5190 -0.04423 ± 0.00891 0.34501 ± 0.01046 13.764 ± 0.298 % 80.232 ± 0.345 %
|
| 20 |
+
14 20.2610 ± 0.5377 -0.03562 ± 0.00840 0.33139 ± 0.00973 13.449 ± 0.284 % 80.052 ± 0.334 %
|
| 21 |
+
15 19.6396 ± 0.5014 -0.03001 ± 0.00796 0.31598 ± 0.00910 13.229 ± 0.271 % 80.391 ± 0.321 %
|
| 22 |
+
16 18.2437 ± 0.4482 -0.02734 ± 0.00755 0.30048 ± 0.00857 12.932 ± 0.261 % 81.103 ± 0.306 %
|
| 23 |
+
17 18.1996 ± 0.4304 -0.02528 ± 0.00719 0.28990 ± 0.00808 12.683 ± 0.251 % 81.111 ± 0.297 %
|
| 24 |
+
18 18.7860 ± 0.4310 -0.02197 ± 0.00685 0.27922 ± 0.00764 12.440 ± 0.242 % 81.139 ± 0.288 %
|
| 25 |
+
19 19.0784 ± 0.4252 -0.01457 ± 0.00659 0.27199 ± 0.00725 12.275 ± 0.233 % 81.062 ± 0.281 %
|
| 26 |
+
20 19.6134 ± 0.4257 -0.00840 ± 0.00634 0.26519 ± 0.00689 12.079 ± 0.226 % 81.061 ± 0.274 %
|
| 27 |
+
21 19.8939 ± 0.4212 -0.00176 ± 0.00610 0.25827 ± 0.00657 11.900 ± 0.219 % 81.120 ± 0.267 %
|
| 28 |
+
22 19.1779 ± 0.3941 -0.00042 ± 0.00585 0.24875 ± 0.00628 11.704 ± 0.213 % 81.565 ± 0.258 %
|
| 29 |
+
23 18.1599 ± 0.3630 -0.00130 ± 0.00562 0.23988 ± 0.00602 11.510 ± 0.207 % 81.992 ± 0.251 %
|
| 30 |
+
24 18.3523 ± 0.3604 0.00148 ± 0.00548 0.23611 ± 0.00578 11.406 ± 0.201 % 81.965 ± 0.245 %
|
| 31 |
+
25 17.6893 ± 0.3389 0.00199 ± 0.00528 0.22815 ± 0.00556 11.223 ± 0.196 % 82.350 ± 0.238 %
|
| 32 |
+
26 18.0398 ± 0.3388 0.00519 ± 0.00512 0.22323 ± 0.00535 11.094 ± 0.191 % 82.326 ± 0.234 %
|
| 33 |
+
27 17.2077 ± 0.3162 0.00591 ± 0.00498 0.21861 ± 0.00516 11.067 ± 0.186 % 82.604 ± 0.228 %
|
| 34 |
+
28 17.4397 ± 0.3143 0.00703 ± 0.00484 0.21471 ± 0.00498 10.962 ± 0.181 % 82.625 ± 0.224 %
|
| 35 |
+
29 17.8198 ± 0.3168 0.01122 ± 0.00474 0.21251 ± 0.00481 10.920 ± 0.177 % 82.502 ± 0.221 %
|
| 36 |
+
30 18.2178 ± 0.3180 0.01087 ± 0.00464 0.21008 ± 0.00466 10.842 ± 0.173 % 82.398 ± 0.217 %
|
| 37 |
+
31 18.0640 ± 0.3089 0.01176 ± 0.00452 0.20629 ± 0.00451 10.772 ± 0.169 % 82.461 ± 0.214 %
|
| 38 |
+
32 17.6394 ± 0.2959 0.00232 ± 0.00456 0.21097 ± 0.00444 11.023 ± 0.166 % 82.310 ± 0.211 %
|
| 39 |
+
33 17.9033 ± 0.2954 0.00318 ± 0.00446 0.20772 ± 0.00431 10.918 ± 0.163 % 82.254 ± 0.208 %
|
| 40 |
+
34 17.9506 ± 0.2909 0.00351 ± 0.00436 0.20459 ± 0.00418 10.853 ± 0.159 % 82.238 ± 0.205 %
|
| 41 |
+
35 18.2135 ± 0.2917 0.00497 ± 0.00428 0.20221 �� 0.00407 10.793 ± 0.156 % 82.265 ± 0.202 %
|
| 42 |
+
36 18.6417 ± 0.2959 0.00195 ± 0.00464 0.24893 ± 0.00457 11.471 ± 0.157 % 81.194 ± 0.204 %
|
| 43 |
+
37 18.8949 ± 0.2959 0.00337 ± 0.00454 0.24459 ± 0.00445 11.358 ± 0.154 % 81.319 ± 0.200 %
|
| 44 |
+
38 19.3157 ± 0.2992 0.00392 ± 0.00444 0.24126 ± 0.00433 11.275 ± 0.152 % 81.275 ± 0.198 %
|
| 45 |
+
39 19.5740 ± 0.2992 0.00539 ± 0.00435 0.23784 ± 0.00422 11.198 ± 0.149 % 81.257 ± 0.195 %
|
| 46 |
+
40 19.7017 ± 0.2971 0.00603 ± 0.00426 0.23443 ± 0.00412 11.108 ± 0.147 % 81.261 ± 0.193 %
|
| 47 |
+
41 19.9681 ± 0.2978 0.00745 ± 0.00418 0.23180 ± 0.00402 11.039 ± 0.144 % 81.189 ± 0.191 %
|
| 48 |
+
42 20.2624 ± 0.2992 0.00920 ± 0.00411 0.22991 ± 0.00393 10.986 ± 0.142 % 81.141 ± 0.189 %
|
| 49 |
+
43 20.5921 ± 0.3007 0.01051 ± 0.00404 0.22778 ± 0.00384 10.918 ± 0.139 % 81.125 ± 0.187 %
|
| 50 |
+
44 20.2269 ± 0.2910 0.01010 ± 0.00396 0.22299 ± 0.00376 10.805 ± 0.138 % 81.387 ± 0.183 %
|
| 51 |
+
|
| 52 |
+
====== Perplexity statistics ======
|
| 53 |
+
Mean PPL(Q) : 20.226883 ± 0.291005
|
| 54 |
+
Mean PPL(base) : 20.023532 ± 0.285215
|
| 55 |
+
Cor(ln(PPL(Q)), ln(PPL(base))): 96.19%
|
| 56 |
+
Mean ln(PPL(Q)/PPL(base)) : 0.010104 ± 0.003955
|
| 57 |
+
Mean PPL(Q)/PPL(base) : 1.010156 ± 0.003995
|
| 58 |
+
Mean PPL(Q)-PPL(base) : 0.203352 ± 0.079759
|
| 59 |
+
|
| 60 |
+
====== KL divergence statistics ======
|
| 61 |
+
Mean KLD: 0.222994 ± 0.003760
|
| 62 |
+
Maximum KLD: 21.454760
|
| 63 |
+
99.9% KLD: 11.217873
|
| 64 |
+
99.0% KLD: 3.457192
|
| 65 |
+
95.0% KLD: 0.742914
|
| 66 |
+
90.0% KLD: 0.354640
|
| 67 |
+
Median KLD: 0.061749
|
| 68 |
+
10.0% KLD: 0.003209
|
| 69 |
+
5.0% KLD: 0.000565
|
| 70 |
+
1.0% KLD: 0.000016
|
| 71 |
+
0.1% KLD: 0.000000
|
| 72 |
+
Minimum KLD: -0.000076
|
| 73 |
+
|
| 74 |
+
====== Token probability statistics ======
|
| 75 |
+
Mean Δp: 0.222 ± 0.051 %
|
| 76 |
+
Maximum Δp: 99.968%
|
| 77 |
+
99.9% Δp: 97.018%
|
| 78 |
+
99.0% Δp: 39.167%
|
| 79 |
+
95.0% Δp: 12.490%
|
| 80 |
+
90.0% Δp: 6.425%
|
| 81 |
+
75.0% Δp: 0.990%
|
| 82 |
+
Median Δp: -0.002%
|
| 83 |
+
25.0% Δp: -1.192%
|
| 84 |
+
10.0% Δp: -6.597%
|
| 85 |
+
5.0% Δp: -12.478%
|
| 86 |
+
1.0% Δp: -31.393%
|
| 87 |
+
0.1% Δp: -65.134%
|
| 88 |
+
Minimum Δp: -91.785%
|
| 89 |
+
RMS Δp : 10.805 ± 0.138 %
|
| 90 |
+
Same top p: 81.387 ± 0.183 %
|
| 91 |
+
|
reports/kld2048b-Q5_K_M.txt
ADDED
|
@@ -0,0 +1,91 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
0.00.692.086 I
|
| 2 |
+
0.00.692.109 I system_info: n_threads = 12 (n_threads_batch = 12) / 16 | MTL : EMBED_LIBRARY = 1 | CPU : NEON = 1 | ARM_FMA = 1 | FP16_VA = 1 | MATMUL_INT8 = 1 | DOTPROD = 1 | SME = 1 | ACCELERATE = 1 | OPENMP = 1 | REPACK = 1 |
|
| 3 |
+
0.00.707.153 I kl_divergence: computing over 44 chunks, n_ctx=2048, batch_size=2048, n_seq=1
|
| 4 |
+
0.05.303.346 I kl_divergence: 4.60 seconds per pass - ETA 3.37 minutes
|
| 5 |
+
|
| 6 |
+
chunk PPL ln(PPL(Q)/PPL(base)) KL Divergence Δp RMS Same top p
|
| 7 |
+
1 39.9630 ± 4.0696 0.01932 ± 0.01341 0.05495 ± 0.00231 5.203 ± 0.376 % 86.217 ± 1.078 %
|
| 8 |
+
2 43.5242 ± 3.1197 0.00601 ± 0.01041 0.06186 ± 0.00216 5.467 ± 0.266 % 85.386 ± 0.781 %
|
| 9 |
+
3 30.9644 ± 1.8103 0.00145 ± 0.00731 0.04755 ± 0.00155 4.832 ± 0.211 % 88.302 ± 0.580 %
|
| 10 |
+
4 21.1607 ± 1.0531 -0.00227 ± 0.00611 0.04406 ± 0.00272 4.902 ± 0.299 % 89.370 ± 0.482 %
|
| 11 |
+
5 16.6963 ± 0.7261 -0.00247 ± 0.00511 0.03802 ± 0.00223 4.682 ± 0.268 % 90.694 ± 0.406 %
|
| 12 |
+
6 18.8603 ± 0.7547 -0.00216 ± 0.00504 0.04566 ± 0.00203 5.261 ± 0.258 % 89.541 ± 0.391 %
|
| 13 |
+
7 21.5013 ± 0.8064 0.00541 ± 0.00476 0.04762 ± 0.00179 5.240 ± 0.228 % 89.317 ± 0.365 %
|
| 14 |
+
8 23.7471 ± 0.8336 0.00652 ± 0.00456 0.05059 ± 0.00162 5.252 ± 0.205 % 88.685 ± 0.350 %
|
| 15 |
+
9 20.8119 ± 0.6849 0.00489 ± 0.00421 0.04720 ± 0.00147 5.094 ± 0.190 % 89.399 ± 0.321 %
|
| 16 |
+
10 23.2395 ± 0.7341 0.00020 ± 0.00401 0.04927 ± 0.00136 5.073 ± 0.176 % 88.915 ± 0.310 %
|
| 17 |
+
11 20.7576 ± 0.6221 -0.00019 ± 0.00379 0.04718 ± 0.00126 5.073 ± 0.166 % 89.514 ± 0.289 %
|
| 18 |
+
12 19.8390 ± 0.5638 -0.00673 ± 0.00362 0.04730 ± 0.00118 5.261 ± 0.159 % 89.589 ± 0.276 %
|
| 19 |
+
13 19.7641 ± 0.5371 -0.00316 ± 0.00347 0.04832 ± 0.00112 5.362 ± 0.152 % 89.338 ± 0.268 %
|
| 20 |
+
14 20.8728 ± 0.5490 -0.00587 ± 0.00341 0.05021 ± 0.00107 5.379 ± 0.142 % 88.807 ± 0.263 %
|
| 21 |
+
15 20.1173 ± 0.5087 -0.00597 ± 0.00331 0.05014 ± 0.00102 5.454 ± 0.134 % 88.895 ± 0.254 %
|
| 22 |
+
16 18.6445 ± 0.4542 -0.00560 ± 0.00314 0.04803 ± 0.00097 5.374 ± 0.129 % 89.266 ± 0.242 %
|
| 23 |
+
17 18.5453 ± 0.4345 -0.00646 ± 0.00302 0.04736 ± 0.00091 5.327 ± 0.123 % 89.265 ± 0.235 %
|
| 24 |
+
18 19.0987 ± 0.4338 -0.00546 ± 0.00289 0.04654 ± 0.00087 5.293 ± 0.118 % 89.302 ± 0.228 %
|
| 25 |
+
19 19.3034 ± 0.4256 -0.00284 ± 0.00282 0.04721 ± 0.00083 5.325 ± 0.113 % 89.165 ± 0.223 %
|
| 26 |
+
20 19.7218 ± 0.4225 -0.00289 ± 0.00276 0.04737 ± 0.00080 5.303 ± 0.109 % 89.115 ± 0.218 %
|
| 27 |
+
21 19.8754 ± 0.4148 -0.00269 ± 0.00267 0.04721 ± 0.00077 5.272 ± 0.105 % 89.168 ± 0.212 %
|
| 28 |
+
22 19.1387 ± 0.3879 -0.00247 ± 0.00257 0.04578 ± 0.00074 5.198 ± 0.102 % 89.429 ± 0.205 %
|
| 29 |
+
23 18.1283 ± 0.3576 -0.00304 ± 0.00248 0.04464 ± 0.00074 5.156 ± 0.101 % 89.715 ± 0.198 %
|
| 30 |
+
24 18.2849 ± 0.3541 -0.00220 ± 0.00245 0.04493 ± 0.00073 5.192 ± 0.098 % 89.622 ± 0.195 %
|
| 31 |
+
25 17.6103 ± 0.3329 -0.00249 ± 0.00236 0.04363 ± 0.00070 5.118 ± 0.095 % 89.822 ± 0.189 %
|
| 32 |
+
26 17.9036 ± 0.3319 -0.00239 ± 0.00232 0.04353 ± 0.00068 5.085 ± 0.093 % 89.785 ± 0.186 %
|
| 33 |
+
27 17.0832 ± 0.3098 -0.00135 ± 0.00229 0.04387 ± 0.00069 5.237 ± 0.096 % 89.870 ± 0.182 %
|
| 34 |
+
28 17.2934 ± 0.3076 -0.00139 ± 0.00224 0.04378 ± 0.00067 5.228 ± 0.094 % 89.862 ± 0.178 %
|
| 35 |
+
29 17.5885 ± 0.3082 -0.00184 ± 0.00222 0.04429 ± 0.00065 5.250 ± 0.091 % 89.770 ± 0.176 %
|
| 36 |
+
30 17.9869 ± 0.3097 -0.00188 ± 0.00219 0.04490 ± 0.00064 5.247 ± 0.089 % 89.642 ± 0.174 %
|
| 37 |
+
31 17.8276 ± 0.3007 -0.00142 ± 0.00215 0.04470 ± 0.00062 5.253 ± 0.088 % 89.686 ± 0.171 %
|
| 38 |
+
32 17.4479 ± 0.2893 -0.00859 ± 0.00233 0.05061 ± 0.00086 5.790 ± 0.106 % 89.421 ± 0.170 %
|
| 39 |
+
33 17.6847 ± 0.2885 -0.00910 ± 0.00229 0.05052 ± 0.00084 5.774 ± 0.103 % 89.369 ± 0.168 %
|
| 40 |
+
34 17.7162 ± 0.2839 -0.00964 ± 0.00225 0.05027 ± 0.00082 5.754 ± 0.101 % 89.293 ± 0.166 %
|
| 41 |
+
35 17.9647 ± 0.2843 -0.00879 ± 0.00222 0.05104 �� 0.00081 5.788 ± 0.099 % 89.180 ± 0.164 %
|
| 42 |
+
36 18.4521 ± 0.2895 -0.00827 ± 0.00224 0.05340 ± 0.00082 5.846 ± 0.096 % 88.938 ± 0.163 %
|
| 43 |
+
37 18.6706 ± 0.2888 -0.00858 ± 0.00220 0.05283 ± 0.00080 5.802 ± 0.094 % 88.999 ± 0.161 %
|
| 44 |
+
38 19.0670 ± 0.2914 -0.00904 ± 0.00216 0.05264 ± 0.00078 5.783 ± 0.092 % 88.946 ± 0.159 %
|
| 45 |
+
39 19.2958 ± 0.2909 -0.00893 ± 0.00212 0.05232 ± 0.00076 5.754 ± 0.091 % 88.937 ± 0.157 %
|
| 46 |
+
40 19.3982 ± 0.2885 -0.00950 ± 0.00208 0.05215 ± 0.00074 5.724 ± 0.089 % 88.917 ± 0.155 %
|
| 47 |
+
41 19.6150 ± 0.2883 -0.01039 ± 0.00205 0.05202 ± 0.00073 5.699 ± 0.087 % 88.813 ± 0.154 %
|
| 48 |
+
42 19.8756 ± 0.2890 -0.01008 ± 0.00203 0.05223 ± 0.00071 5.695 ± 0.086 % 88.773 ± 0.152 %
|
| 49 |
+
43 20.1426 ± 0.2895 -0.01156 ± 0.00200 0.05217 ± 0.00070 5.676 ± 0.084 % 88.743 ± 0.151 %
|
| 50 |
+
44 19.7970 ± 0.2804 -0.01138 ± 0.00196 0.05113 ± 0.00068 5.620 ± 0.083 % 88.887 ± 0.148 %
|
| 51 |
+
|
| 52 |
+
====== Perplexity statistics ======
|
| 53 |
+
Mean PPL(Q) : 19.797018 ± 0.280375
|
| 54 |
+
Mean PPL(base) : 20.023532 ± 0.285215
|
| 55 |
+
Cor(ln(PPL(Q)), ln(PPL(base))): 99.05%
|
| 56 |
+
Mean ln(PPL(Q)/PPL(base)) : -0.011377 ± 0.001959
|
| 57 |
+
Mean PPL(Q)/PPL(base) : 0.988688 ± 0.001937
|
| 58 |
+
Mean PPL(Q)-PPL(base) : -0.226513 ± 0.039268
|
| 59 |
+
|
| 60 |
+
====== KL divergence statistics ======
|
| 61 |
+
Mean KLD: 0.051125 ± 0.000685
|
| 62 |
+
Maximum KLD: 9.869919
|
| 63 |
+
99.9% KLD: 1.489997
|
| 64 |
+
99.0% KLD: 0.466226
|
| 65 |
+
95.0% KLD: 0.174651
|
| 66 |
+
90.0% KLD: 0.109054
|
| 67 |
+
Median KLD: 0.022373
|
| 68 |
+
10.0% KLD: 0.001169
|
| 69 |
+
5.0% KLD: 0.000194
|
| 70 |
+
1.0% KLD: 0.000005
|
| 71 |
+
0.1% KLD: -0.000001
|
| 72 |
+
Minimum KLD: -0.000079
|
| 73 |
+
|
| 74 |
+
====== Token probability statistics ======
|
| 75 |
+
Mean Δp: -0.136 ± 0.026 %
|
| 76 |
+
Maximum Δp: 99.862%
|
| 77 |
+
99.9% Δp: 44.541%
|
| 78 |
+
99.0% Δp: 17.103%
|
| 79 |
+
95.0% Δp: 6.634%
|
| 80 |
+
90.0% Δp: 3.349%
|
| 81 |
+
75.0% Δp: 0.531%
|
| 82 |
+
Median Δp: -0.000%
|
| 83 |
+
25.0% Δp: -0.740%
|
| 84 |
+
10.0% Δp: -4.072%
|
| 85 |
+
5.0% Δp: -7.444%
|
| 86 |
+
1.0% Δp: -18.135%
|
| 87 |
+
0.1% Δp: -39.395%
|
| 88 |
+
Minimum Δp: -82.381%
|
| 89 |
+
RMS Δp : 5.620 ± 0.083 %
|
| 90 |
+
Same top p: 88.887 ± 0.148 %
|
| 91 |
+
|
reports/kld2048b-Q6_K.txt
ADDED
|
@@ -0,0 +1,91 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
0.00.791.202 I
|
| 2 |
+
0.00.791.226 I system_info: n_threads = 12 (n_threads_batch = 12) / 16 | MTL : EMBED_LIBRARY = 1 | CPU : NEON = 1 | ARM_FMA = 1 | FP16_VA = 1 | MATMUL_INT8 = 1 | DOTPROD = 1 | SME = 1 | ACCELERATE = 1 | OPENMP = 1 | REPACK = 1 |
|
| 3 |
+
0.00.806.466 I kl_divergence: computing over 44 chunks, n_ctx=2048, batch_size=2048, n_seq=1
|
| 4 |
+
0.06.581.283 I kl_divergence: 5.77 seconds per pass - ETA 4.23 minutes
|
| 5 |
+
|
| 6 |
+
chunk PPL ln(PPL(Q)/PPL(base)) KL Divergence Δp RMS Same top p
|
| 7 |
+
1 39.7693 ± 4.0512 0.01446 ± 0.01296 0.04919 ± 0.00219 4.913 ± 0.344 % 85.826 ± 1.091 %
|
| 8 |
+
2 42.9960 ± 3.0837 -0.00621 ± 0.00972 0.05312 ± 0.00169 4.983 ± 0.248 % 86.315 ± 0.760 %
|
| 9 |
+
3 30.7245 ± 1.7950 -0.00633 ± 0.00679 0.03963 ± 0.00120 4.331 ± 0.194 % 89.084 ± 0.563 %
|
| 10 |
+
4 21.0200 ± 1.0464 -0.00894 ± 0.00557 0.03574 ± 0.00228 4.472 ± 0.310 % 90.591 ± 0.456 %
|
| 11 |
+
5 16.6222 ± 0.7230 -0.00692 ± 0.00456 0.02987 ± 0.00184 4.137 ± 0.270 % 91.945 ± 0.381 %
|
| 12 |
+
6 18.8735 ± 0.7573 -0.00146 ± 0.00443 0.03418 ± 0.00161 4.531 ± 0.235 % 91.088 ± 0.364 %
|
| 13 |
+
7 21.3666 ± 0.8029 -0.00087 ± 0.00432 0.03786 ± 0.00147 4.667 ± 0.211 % 90.476 ± 0.347 %
|
| 14 |
+
8 23.5987 ± 0.8302 0.00025 ± 0.00420 0.04052 ± 0.00133 4.676 ± 0.187 % 89.907 ± 0.333 %
|
| 15 |
+
9 20.6999 ± 0.6823 -0.00051 ± 0.00384 0.03759 ± 0.00119 4.502 ± 0.173 % 90.453 ± 0.306 %
|
| 16 |
+
10 23.0941 ± 0.7301 -0.00607 ± 0.00366 0.03977 ± 0.00111 4.498 ± 0.160 % 89.814 ± 0.299 %
|
| 17 |
+
11 20.6333 ± 0.6188 -0.00620 ± 0.00341 0.03763 ± 0.00102 4.404 ± 0.149 % 90.411 ± 0.278 %
|
| 18 |
+
12 19.9340 ± 0.5679 -0.00194 ± 0.00335 0.04076 ± 0.00099 4.744 ± 0.143 % 90.176 ± 0.269 %
|
| 19 |
+
13 19.8184 ± 0.5400 -0.00042 ± 0.00320 0.04112 ± 0.00093 4.836 ± 0.136 % 90.105 ± 0.259 %
|
| 20 |
+
14 20.8748 ± 0.5505 -0.00578 ± 0.00316 0.04311 ± 0.00090 4.876 ± 0.128 % 89.652 ± 0.255 %
|
| 21 |
+
15 20.1217 ± 0.5101 -0.00575 ± 0.00307 0.04303 ± 0.00085 4.975 ± 0.121 % 89.697 ± 0.245 %
|
| 22 |
+
16 18.6525 ± 0.4555 -0.00517 ± 0.00290 0.04100 ± 0.00081 4.884 ± 0.116 % 90.090 ± 0.234 %
|
| 23 |
+
17 18.5724 ± 0.4363 -0.00500 ± 0.00279 0.04037 ± 0.00077 4.825 ± 0.111 % 90.087 ± 0.227 %
|
| 24 |
+
18 19.0767 ± 0.4345 -0.00661 ± 0.00268 0.03954 ± 0.00073 4.780 ± 0.107 % 90.187 ± 0.219 %
|
| 25 |
+
19 19.2489 ± 0.4255 -0.00567 ± 0.00261 0.03969 ± 0.00070 4.786 ± 0.102 % 90.096 ± 0.214 %
|
| 26 |
+
20 19.6439 ± 0.4222 -0.00685 ± 0.00254 0.03964 ± 0.00067 4.754 ± 0.098 % 90.049 ± 0.209 %
|
| 27 |
+
21 19.7655 ± 0.4138 -0.00824 ± 0.00246 0.03914 ± 0.00064 4.718 ± 0.095 % 90.113 ± 0.204 %
|
| 28 |
+
22 19.0431 ± 0.3872 -0.00748 ± 0.00236 0.03785 ± 0.00061 4.647 ± 0.092 % 90.354 ± 0.197 %
|
| 29 |
+
23 18.0370 ± 0.3568 -0.00809 ± 0.00227 0.03679 ± 0.00060 4.609 ± 0.092 % 90.565 ± 0.191 %
|
| 30 |
+
24 18.1912 ± 0.3532 -0.00733 ± 0.00225 0.03715 ± 0.00059 4.656 ± 0.090 % 90.441 ± 0.188 %
|
| 31 |
+
25 17.5234 ± 0.3320 -0.00744 ± 0.00217 0.03602 ± 0.00056 4.592 ± 0.088 % 90.663 ± 0.182 %
|
| 32 |
+
26 17.7978 ± 0.3305 -0.00831 ± 0.00212 0.03589 ± 0.00054 4.584 ± 0.086 % 90.627 ± 0.179 %
|
| 33 |
+
27 16.9643 ± 0.3080 -0.00833 ± 0.00209 0.03595 ± 0.00055 4.681 ± 0.088 % 90.743 ± 0.174 %
|
| 34 |
+
28 17.1752 ± 0.3058 -0.00825 ± 0.00204 0.03580 ± 0.00053 4.683 ± 0.085 % 90.738 ± 0.171 %
|
| 35 |
+
29 17.4899 ± 0.3070 -0.00746 ± 0.00202 0.03632 ± 0.00052 4.704 ± 0.083 % 90.633 ± 0.169 %
|
| 36 |
+
30 17.8846 ± 0.3083 -0.00759 ± 0.00200 0.03698 ± 0.00051 4.713 ± 0.080 % 90.505 ± 0.167 %
|
| 37 |
+
31 17.7369 ± 0.2995 -0.00652 ± 0.00197 0.03677 ± 0.00050 4.729 ± 0.080 % 90.540 ± 0.164 %
|
| 38 |
+
32 17.5366 ± 0.2916 -0.00352 ± 0.00196 0.03685 ± 0.00049 4.777 ± 0.079 % 90.558 ± 0.162 %
|
| 39 |
+
33 17.7734 ± 0.2908 -0.00410 ± 0.00193 0.03695 ± 0.00048 4.767 ± 0.077 % 90.471 ± 0.160 %
|
| 40 |
+
34 17.8213 ± 0.2865 -0.00372 ± 0.00189 0.03687 ± 0.00047 4.763 ± 0.076 % 90.363 ± 0.158 %
|
| 41 |
+
35 18.0669 ± 0.2869 -0.00312 ± 0.00187 0.03753 �� 0.00048 4.800 ± 0.075 % 90.292 ± 0.156 %
|
| 42 |
+
36 18.5795 ± 0.2929 -0.00139 ± 0.00278 0.08702 ± 0.00233 6.398 ± 0.131 % 89.011 ± 0.163 %
|
| 43 |
+
37 18.8027 ± 0.2923 -0.00152 ± 0.00271 0.08538 ± 0.00227 6.337 ± 0.129 % 89.118 ± 0.160 %
|
| 44 |
+
38 19.1927 ± 0.2947 -0.00247 ± 0.00265 0.08427 ± 0.00221 6.297 ± 0.127 % 89.036 ± 0.158 %
|
| 45 |
+
39 19.4273 ± 0.2944 -0.00214 ± 0.00260 0.08312 ± 0.00216 6.252 ± 0.124 % 88.999 ± 0.157 %
|
| 46 |
+
40 19.5252 ± 0.2918 -0.00297 ± 0.00255 0.08206 ± 0.00210 6.210 ± 0.122 % 88.969 ± 0.155 %
|
| 47 |
+
41 19.7581 ± 0.2919 -0.00312 ± 0.00250 0.08099 ± 0.00205 6.170 ± 0.120 % 88.964 ± 0.153 %
|
| 48 |
+
42 20.0080 ± 0.2924 -0.00344 ± 0.00246 0.08034 ± 0.00200 6.150 ± 0.118 % 88.905 ± 0.152 %
|
| 49 |
+
43 20.2833 ± 0.2930 -0.00460 ± 0.00242 0.07954 ± 0.00196 6.114 ± 0.116 % 88.895 ± 0.150 %
|
| 50 |
+
44 19.9327 ± 0.2838 -0.00455 ± 0.00236 0.07786 ± 0.00192 6.050 ± 0.114 % 89.056 ± 0.147 %
|
| 51 |
+
|
| 52 |
+
====== Perplexity statistics ======
|
| 53 |
+
Mean PPL(Q) : 19.932671 ± 0.283759
|
| 54 |
+
Mean PPL(base) : 20.023532 ± 0.285215
|
| 55 |
+
Cor(ln(PPL(Q)), ln(PPL(base))): 98.62%
|
| 56 |
+
Mean ln(PPL(Q)/PPL(base)) : -0.004548 ± 0.002363
|
| 57 |
+
Mean PPL(Q)/PPL(base) : 0.995462 ± 0.002352
|
| 58 |
+
Mean PPL(Q)-PPL(base) : -0.090860 ± 0.047230
|
| 59 |
+
|
| 60 |
+
====== KL divergence statistics ======
|
| 61 |
+
Mean KLD: 0.077855 ± 0.001915
|
| 62 |
+
Maximum KLD: 16.764179
|
| 63 |
+
99.9% KLD: 6.203948
|
| 64 |
+
99.0% KLD: 1.520848
|
| 65 |
+
95.0% KLD: 0.174014
|
| 66 |
+
90.0% KLD: 0.098078
|
| 67 |
+
Median KLD: 0.018081
|
| 68 |
+
10.0% KLD: 0.000850
|
| 69 |
+
5.0% KLD: 0.000131
|
| 70 |
+
1.0% KLD: 0.000003
|
| 71 |
+
0.1% KLD: -0.000001
|
| 72 |
+
Minimum KLD: -0.000081
|
| 73 |
+
|
| 74 |
+
====== Token probability statistics ======
|
| 75 |
+
Mean Δp: 0.020 ± 0.029 %
|
| 76 |
+
Maximum Δp: 99.854%
|
| 77 |
+
99.9% Δp: 62.315%
|
| 78 |
+
99.0% Δp: 16.727%
|
| 79 |
+
95.0% Δp: 6.435%
|
| 80 |
+
90.0% Δp: 3.405%
|
| 81 |
+
75.0% Δp: 0.568%
|
| 82 |
+
Median Δp: 0.000%
|
| 83 |
+
25.0% Δp: -0.562%
|
| 84 |
+
10.0% Δp: -3.505%
|
| 85 |
+
5.0% Δp: -6.514%
|
| 86 |
+
1.0% Δp: -16.873%
|
| 87 |
+
0.1% Δp: -49.872%
|
| 88 |
+
Minimum Δp: -97.627%
|
| 89 |
+
RMS Δp : 6.050 ± 0.114 %
|
| 90 |
+
Same top p: 89.056 ± 0.147 %
|
| 91 |
+
|
reports/kld2048b-Q8_0.txt
ADDED
|
@@ -0,0 +1,91 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
0.03.902.859 I
|
| 2 |
+
0.03.902.882 I system_info: n_threads = 12 (n_threads_batch = 12) / 16 | MTL : EMBED_LIBRARY = 1 | CPU : NEON = 1 | ARM_FMA = 1 | FP16_VA = 1 | MATMUL_INT8 = 1 | DOTPROD = 1 | SME = 1 | ACCELERATE = 1 | OPENMP = 1 | REPACK = 1 |
|
| 3 |
+
0.03.917.638 I kl_divergence: computing over 44 chunks, n_ctx=2048, batch_size=2048, n_seq=1
|
| 4 |
+
0.06.676.459 I kl_divergence: 2.76 seconds per pass - ETA 2.02 minutes
|
| 5 |
+
|
| 6 |
+
chunk PPL ln(PPL(Q)/PPL(base)) KL Divergence Δp RMS Same top p
|
| 7 |
+
1 39.3179 ± 3.9687 0.00304 ± 0.00447 0.00512 ± 0.00020 1.719 ± 0.116 % 95.992 ± 0.614 %
|
| 8 |
+
2 43.7425 ± 3.1358 0.01101 ± 0.00482 0.01071 ± 0.00056 2.402 ± 0.135 % 94.379 ± 0.509 %
|
| 9 |
+
3 31.1566 ± 1.8236 0.00764 ± 0.00333 0.00775 ± 0.00040 2.043 ± 0.108 % 95.797 ± 0.362 %
|
| 10 |
+
4 21.3343 ± 1.0652 0.00590 ± 0.00256 0.00627 ± 0.00030 1.885 ± 0.090 % 96.359 ± 0.293 %
|
| 11 |
+
5 16.8234 ± 0.7337 0.00512 ± 0.00210 0.00519 ± 0.00025 1.726 ± 0.079 % 96.950 ± 0.240 %
|
| 12 |
+
6 19.0176 ± 0.7650 0.00615 ± 0.00192 0.00556 ± 0.00022 1.787 ± 0.071 % 96.693 ± 0.228 %
|
| 13 |
+
7 21.5324 ± 0.8110 0.00686 ± 0.00177 0.00567 ± 0.00019 1.783 ± 0.064 % 96.411 ± 0.220 %
|
| 14 |
+
8 23.7777 ± 0.8379 0.00781 ± 0.00163 0.00568 ± 0.00017 1.738 ± 0.058 % 96.261 ± 0.210 %
|
| 15 |
+
9 20.8430 ± 0.6883 0.00638 ± 0.00148 0.00523 ± 0.00015 1.671 ± 0.054 % 96.470 ± 0.192 %
|
| 16 |
+
10 23.4044 ± 0.7421 0.00727 ± 0.00140 0.00541 ± 0.00016 1.658 ± 0.050 % 96.325 ± 0.186 %
|
| 17 |
+
11 20.9095 ± 0.6296 0.00710 ± 0.00132 0.00510 ± 0.00015 1.624 ± 0.047 % 96.561 ± 0.172 %
|
| 18 |
+
12 20.1019 ± 0.5738 0.00644 ± 0.00124 0.00504 ± 0.00014 1.643 ± 0.045 % 96.562 ± 0.164 %
|
| 19 |
+
13 19.9460 ± 0.5443 0.00600 ± 0.00118 0.00507 ± 0.00013 1.675 ± 0.042 % 96.534 ± 0.159 %
|
| 20 |
+
14 21.1268 ± 0.5585 0.00622 ± 0.00116 0.00532 ± 0.00013 1.688 ± 0.040 % 96.327 ± 0.157 %
|
| 21 |
+
15 20.3644 ± 0.5177 0.00623 ± 0.00111 0.00519 ± 0.00012 1.697 ± 0.038 % 96.357 ± 0.151 %
|
| 22 |
+
16 18.8566 ± 0.4618 0.00571 ± 0.00105 0.00494 ± 0.00011 1.667 ± 0.037 % 96.518 ± 0.143 %
|
| 23 |
+
17 18.7770 ± 0.4422 0.00596 ± 0.00101 0.00486 ± 0.00011 1.653 ± 0.035 % 96.527 ± 0.139 %
|
| 24 |
+
18 19.3249 ± 0.4413 0.00632 ± 0.00096 0.00475 ± 0.00010 1.634 ± 0.034 % 96.573 ± 0.134 %
|
| 25 |
+
19 19.4971 ± 0.4323 0.00714 ± 0.00093 0.00471 ± 0.00010 1.629 ± 0.032 % 96.584 ± 0.130 %
|
| 26 |
+
20 19.9216 ± 0.4295 0.00719 ± 0.00090 0.00470 ± 0.00009 1.623 ± 0.031 % 96.608 ± 0.127 %
|
| 27 |
+
21 20.0633 ± 0.4215 0.00672 ± 0.00087 0.00462 ± 0.00009 1.614 ± 0.031 % 96.639 ± 0.123 %
|
| 28 |
+
22 19.3105 ± 0.3940 0.00646 ± 0.00084 0.00448 ± 0.00008 1.589 ± 0.030 % 96.716 ± 0.119 %
|
| 29 |
+
23 18.2955 ± 0.3633 0.00614 ± 0.00080 0.00433 ± 0.00008 1.565 ± 0.029 % 96.791 ± 0.115 %
|
| 30 |
+
24 18.4489 ± 0.3596 0.00673 ± 0.00081 0.00434 ± 0.00008 1.577 ± 0.028 % 96.790 ± 0.112 %
|
| 31 |
+
25 17.7665 ± 0.3379 0.00634 ± 0.00078 0.00421 ± 0.00007 1.556 ± 0.028 % 96.876 ± 0.109 %
|
| 32 |
+
26 18.0634 ± 0.3369 0.00650 ± 0.00076 0.00423 ± 0.00007 1.570 ± 0.027 % 96.857 ± 0.107 %
|
| 33 |
+
27 17.2203 ± 0.3140 0.00664 ± 0.00076 0.00431 ± 0.00008 1.617 ± 0.031 % 96.854 ± 0.105 %
|
| 34 |
+
28 17.4329 ± 0.3118 0.00664 ± 0.00074 0.00427 ± 0.00007 1.609 ± 0.030 % 96.837 ± 0.103 %
|
| 35 |
+
29 17.7293 ± 0.3125 0.00613 ± 0.00074 0.00435 ± 0.00007 1.622 ± 0.029 % 96.794 ± 0.102 %
|
| 36 |
+
30 18.1300 ± 0.3139 0.00604 ± 0.00073 0.00440 ± 0.00007 1.617 ± 0.028 % 96.774 ± 0.101 %
|
| 37 |
+
31 17.9560 ± 0.3043 0.00576 ± 0.00071 0.00436 ± 0.00007 1.617 ± 0.028 % 96.806 ± 0.099 %
|
| 38 |
+
32 17.7479 ± 0.2960 0.00845 ± 0.00077 0.00503 ± 0.00009 1.865 ± 0.047 % 96.722 ± 0.098 %
|
| 39 |
+
33 18.0002 ± 0.2953 0.00858 ± 0.00075 0.00500 ± 0.00009 1.857 ± 0.046 % 96.673 ± 0.098 %
|
| 40 |
+
34 18.0432 ± 0.2910 0.00865 ± 0.00074 0.00496 ± 0.00008 1.848 ± 0.045 % 96.651 ± 0.096 %
|
| 41 |
+
35 18.2803 ± 0.2911 0.00863 ± 0.00073 0.00500 �� 0.00008 1.848 ± 0.044 % 96.612 ± 0.096 %
|
| 42 |
+
36 18.7684 ± 0.2961 0.00872 ± 0.00079 0.00611 ± 0.00011 1.952 ± 0.044 % 96.416 ± 0.097 %
|
| 43 |
+
37 18.9962 ± 0.2956 0.00872 ± 0.00078 0.00603 ± 0.00011 1.934 ± 0.043 % 96.418 ± 0.096 %
|
| 44 |
+
38 19.4098 ± 0.2985 0.00878 ± 0.00076 0.00599 ± 0.00011 1.922 ± 0.042 % 96.393 ± 0.095 %
|
| 45 |
+
39 19.6427 ± 0.2980 0.00889 ± 0.00075 0.00593 ± 0.00011 1.911 ± 0.042 % 96.393 ± 0.093 %
|
| 46 |
+
40 19.7654 ± 0.2958 0.00925 ± 0.00073 0.00590 ± 0.00010 1.904 ± 0.041 % 96.378 ± 0.092 %
|
| 47 |
+
41 20.0024 ± 0.2960 0.00917 ± 0.00072 0.00587 ± 0.00010 1.894 ± 0.040 % 96.366 ± 0.091 %
|
| 48 |
+
42 20.2606 ± 0.2967 0.00911 ± 0.00071 0.00585 ± 0.00010 1.887 ± 0.039 % 96.348 ± 0.090 %
|
| 49 |
+
43 20.5612 ± 0.2977 0.00900 ± 0.00070 0.00583 ± 0.00010 1.879 ± 0.038 % 96.329 ± 0.090 %
|
| 50 |
+
44 20.2006 ± 0.2882 0.00880 ± 0.00069 0.00571 ± 0.00009 1.860 ± 0.038 % 96.381 ± 0.088 %
|
| 51 |
+
|
| 52 |
+
====== Perplexity statistics ======
|
| 53 |
+
Mean PPL(Q) : 20.200587 ± 0.288209
|
| 54 |
+
Mean PPL(base) : 20.023532 ± 0.285215
|
| 55 |
+
Cor(ln(PPL(Q)), ln(PPL(base))): 99.88%
|
| 56 |
+
Mean ln(PPL(Q)/PPL(base)) : 0.008803 ± 0.000686
|
| 57 |
+
Mean PPL(Q)/PPL(base) : 1.008842 ± 0.000692
|
| 58 |
+
Mean PPL(Q)-PPL(base) : 0.177055 ± 0.014106
|
| 59 |
+
|
| 60 |
+
====== KL divergence statistics ======
|
| 61 |
+
Mean KLD: 0.005708 ± 0.000095
|
| 62 |
+
Maximum KLD: 1.475488
|
| 63 |
+
99.9% KLD: 0.233686
|
| 64 |
+
99.0% KLD: 0.060422
|
| 65 |
+
95.0% KLD: 0.019779
|
| 66 |
+
90.0% KLD: 0.011459
|
| 67 |
+
Median KLD: 0.002024
|
| 68 |
+
10.0% KLD: 0.000093
|
| 69 |
+
5.0% KLD: 0.000014
|
| 70 |
+
1.0% KLD: -0.000000
|
| 71 |
+
0.1% KLD: -0.000027
|
| 72 |
+
Minimum KLD: -0.000139
|
| 73 |
+
|
| 74 |
+
====== Token probability statistics ======
|
| 75 |
+
Mean Δp: -0.129 ± 0.009 %
|
| 76 |
+
Maximum Δp: 26.548%
|
| 77 |
+
99.9% Δp: 12.262%
|
| 78 |
+
99.0% Δp: 4.915%
|
| 79 |
+
95.0% Δp: 1.852%
|
| 80 |
+
90.0% Δp: 0.953%
|
| 81 |
+
75.0% Δp: 0.123%
|
| 82 |
+
Median Δp: -0.001%
|
| 83 |
+
25.0% Δp: -0.252%
|
| 84 |
+
10.0% Δp: -1.292%
|
| 85 |
+
5.0% Δp: -2.469%
|
| 86 |
+
1.0% Δp: -6.287%
|
| 87 |
+
0.1% Δp: -15.758%
|
| 88 |
+
Minimum Δp: -56.803%
|
| 89 |
+
RMS Δp : 1.860 ± 0.038 %
|
| 90 |
+
Same top p: 96.381 ± 0.088 %
|
| 91 |
+
|
scripts/run_amalia_pt.sh
ADDED
|
@@ -0,0 +1,88 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/bin/bash
|
| 2 |
+
# Salamandra-7b-instruct full-ladder quant + KLD, calibrated on PORTUGUESE (the AMALIA lesson).
|
| 3 |
+
set -o pipefail
|
| 4 |
+
CONVERT=/Users/kikocisneros/llama.cpp/convert_hf_to_gguf.py
|
| 5 |
+
QUANT=/opt/homebrew/bin/llama-quantize
|
| 6 |
+
IMAT=/opt/homebrew/bin/llama-imatrix
|
| 7 |
+
GGUF=/opt/homebrew/bin/llama-gguf
|
| 8 |
+
PPX=/opt/homebrew/bin/llama-perplexity
|
| 9 |
+
PY=/Users/kikocisneros/work2026/agentic-install/lab/gemma-e2b-cli/.venv/bin/python3
|
| 10 |
+
GL=~/.gpu_coordination/gpu-lock.sh
|
| 11 |
+
W=/Users/kikocisneros/coco_ppl/aq_amalia
|
| 12 |
+
HF_DIR=$W/hf
|
| 13 |
+
F16=$W/amalia-9b-f16.gguf
|
| 14 |
+
IMATRIX=$W/amalia-9b-es.imatrix
|
| 15 |
+
CORPUS=$W/pt_corpus.txt # PORTUGUESE — for imatrix AND KLD eval
|
| 16 |
+
KBASE=$W/kld_base.dat
|
| 17 |
+
REP=$W/reports; mkdir -p "$HF_DIR" "$REP" "$W/metrics"
|
| 18 |
+
LOG=$W/pipeline.log
|
| 19 |
+
exec > >(tee -a "$LOG") 2>&1
|
| 20 |
+
echo "==================== [$(date)] AMALIA PIPELINE START ===================="
|
| 21 |
+
fail(){ echo "!!! FAIL: $1"; bash $GL release agentic-install 2>/dev/null; exit 1; }
|
| 22 |
+
|
| 23 |
+
# 1. DOWNLOAD
|
| 24 |
+
if [ ! -f "$HF_DIR/config.json" ]; then
|
| 25 |
+
echo "--- [1] download ---"
|
| 26 |
+
$PY - <<PY2 || fail "download"
|
| 27 |
+
from huggingface_hub import snapshot_download
|
| 28 |
+
snapshot_download("amalia-llm/AMALIA-9B-0626-SFT", local_dir="$HF_DIR",
|
| 29 |
+
allow_patterns=["*.safetensors","*.json","tokenizer*","*.txt","*.model","*.jinja"], max_workers=4)
|
| 30 |
+
print("DL_DONE")
|
| 31 |
+
PY2
|
| 32 |
+
else echo "--- [1] download present ---"; fi
|
| 33 |
+
|
| 34 |
+
# 2. CONVERT -> F16 (magic-check gates incomplete converts; no fragile load-test)
|
| 35 |
+
if [ ! -f "$F16" ]; then
|
| 36 |
+
echo "--- [2] convert -> F16 ---"
|
| 37 |
+
$PY "$CONVERT" "$HF_DIR" --outfile "$F16" --outtype f16 || fail "convert"
|
| 38 |
+
else echo "--- [2] F16 present ---"; fi
|
| 39 |
+
"$GGUF" "$F16" r 2>&1 | grep -qi "invalid magic" && fail "F16 invalid magic"
|
| 40 |
+
echo "--- [3] load-test SKIPPED (magic verified) ---"
|
| 41 |
+
echo "F16 size: $(du -h "$F16" | cut -f1)"
|
| 42 |
+
|
| 43 |
+
# 4. IMATRIX on PORTUGUESE (GPU)
|
| 44 |
+
if [ ! -f "$IMATRIX" ]; then
|
| 45 |
+
echo "--- [4] imatrix (PORTUGUESE corpus, GPU) ---"
|
| 46 |
+
bash $GL acquire agentic-install high "AMALIA PT imatrix" 2>&1 | tail -1
|
| 47 |
+
"$IMAT" -m "$F16" -f "$CORPUS" -o "$IMATRIX" -c 512 -ngl 99 2>&1 | tail -4 || \
|
| 48 |
+
{ echo "imatrix -ngl 99 fail, retry -ngl 0"; "$IMAT" -m "$F16" -f "$CORPUS" -o "$IMATRIX" -c 512 -ngl 0 2>&1 | tail -4; }
|
| 49 |
+
bash $GL release agentic-install 2>&1 | tail -1
|
| 50 |
+
[ -f "$IMATRIX" ] || fail "imatrix not produced"
|
| 51 |
+
else echo "--- [4] imatrix present ---"; fi
|
| 52 |
+
|
| 53 |
+
# 5. FULL LADDER
|
| 54 |
+
echo "--- [5] full ladder ---"
|
| 55 |
+
for Q in Q3_K_M IQ4_XS Q4_K_M Q5_K_M Q6_K Q8_0; do
|
| 56 |
+
OUTF="$W/amalia-9b-${Q}.gguf"; [ -f "$OUTF" ] && { echo " $Q present"; continue; }
|
| 57 |
+
IMFLAG=""; [ "$Q" != "Q8_0" ] && IMFLAG="--imatrix $IMATRIX"
|
| 58 |
+
"$QUANT" $IMFLAG "$F16" "$OUTF" "$Q" > "$W/q_${Q}.log" 2>&1 || { echo " $Q FAILED"; continue; }
|
| 59 |
+
"$GGUF" "$OUTF" r 2>&1 | grep -qi "invalid magic" && { echo " $Q INVALID MAGIC"; rm -f "$OUTF"; } || echo " $Q OK: $(du -h "$OUTF" | cut -f1)"
|
| 60 |
+
done
|
| 61 |
+
|
| 62 |
+
# 6. KLD SWEEP on PORTUGUESE
|
| 63 |
+
echo "--- [6] KLD sweep (PORTUGUESE) ---"
|
| 64 |
+
bash $GL acquire agentic-install high "AMALIA PT KLD" 2>&1 | tail -1
|
| 65 |
+
"$PPX" -m "$F16" -f "$CORPUS" -c 512 --kl-divergence-base "$KBASE" -ngl 99 > "$REP/base.txt" 2>&1
|
| 66 |
+
F16PPL=$(grep "Final estimate: PPL" "$REP/base.txt" | grep -oE "[0-9]+\.[0-9]+" | head -1)
|
| 67 |
+
CSV=$W/metrics/quant-summary.csv
|
| 68 |
+
echo "name,size_gb,ppl,ppl_delta_vs_f16,kld_mean,kld_p95,kld_max,top1_match_pct" > "$CSV"
|
| 69 |
+
echo "F16,$(echo "scale=2;$(du -m "$F16"|cut -f1)/1024"|bc),${F16PPL},0.0,0.0,0.0,0.0,100.0" >> "$CSV"
|
| 70 |
+
for Q in Q8_0 Q6_K Q5_K_M Q4_K_M IQ4_XS Q3_K_M; do
|
| 71 |
+
M=$W/amalia-9b-${Q}.gguf; [ -f "$M" ] || continue
|
| 72 |
+
R=$REP/kld-${Q}.txt
|
| 73 |
+
"$PPX" -m "$M" -f "$CORPUS" -c 512 --kl-divergence-base "$KBASE" --kl-divergence -ngl 99 > "$R" 2>&1
|
| 74 |
+
PPL=$(grep "Mean PPL(Q) " "$R" | grep -oE "[0-9]+\.[0-9]+" | head -1)
|
| 75 |
+
DELTA=$(grep "Mean PPL(Q)-PPL(base)" "$R" | grep -oE "\-?[0-9]+\.[0-9]+" | head -1)
|
| 76 |
+
KM=$(grep "Mean KLD:" "$R" | grep -oE "[0-9]+\.[0-9]+" | head -1)
|
| 77 |
+
KP95=$(grep "95.0% KLD:" "$R" | sed 's/.*KLD://' | grep -oE "[0-9]+\.[0-9]+" | head -1)
|
| 78 |
+
KMAX=$(grep "Maximum KLD:" "$R" | grep -oE "[0-9]+\.[0-9]+" | head -1)
|
| 79 |
+
T1=$(grep "Same top p:" "$R" | grep -oE "[0-9]+\.[0-9]+" | head -1)
|
| 80 |
+
SZ=$(echo "scale=2;$(du -m "$M"|cut -f1)/1024"|bc)
|
| 81 |
+
echo " $Q: ${SZ}GB ppl=${PPL} dppl=${DELTA} kld=${KM} p95=${KP95} top1=${T1}%"
|
| 82 |
+
echo "${Q},${SZ},${PPL},${DELTA},${KM},${KP95},${KMAX},${T1}" >> "$CSV"
|
| 83 |
+
done
|
| 84 |
+
bash $GL release agentic-install 2>&1 | tail -1
|
| 85 |
+
echo "==================== [$(date)] DONE ===================="
|
| 86 |
+
echo "Produced:"; /bin/ls -lS "$W"/amalia-9b-*.gguf 2>/dev/null | awk '{print " "$5" "$9}'
|
| 87 |
+
echo "--- metrics ---"; cat "$CSV"
|
| 88 |
+
echo "AMALIA_COMPLETE"
|
scripts/run_kld2048.sh
ADDED
|
@@ -0,0 +1,52 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/bin/bash
|
| 2 |
+
# AMALIA-9B — cleaner KLD re-eval at ctx 2048 (lower per-token surprise -> cleaner, monotonic KLD).
|
| 3 |
+
# Reuses the existing F16 + ladder + imatrix; only the KLD eval re-runs. Also a coherence sanity gen.
|
| 4 |
+
set -o pipefail
|
| 5 |
+
QUANT=/opt/homebrew/bin/llama-quantize
|
| 6 |
+
GGUF=/opt/homebrew/bin/llama-gguf
|
| 7 |
+
PPX=/opt/homebrew/bin/llama-perplexity
|
| 8 |
+
CLI=/opt/homebrew/bin/llama-cli
|
| 9 |
+
GL=~/.gpu_coordination/gpu-lock.sh
|
| 10 |
+
W=/Users/kikocisneros/coco_ppl/aq_amalia
|
| 11 |
+
F16=$W/amalia-9b-f16.gguf
|
| 12 |
+
CORPUS=$W/pt_corpus.txt
|
| 13 |
+
KBASE=$W/kld_base_2048.dat
|
| 14 |
+
CSV=$W/metrics/quant-summary.csv # overwrite with the clean ctx-2048 numbers
|
| 15 |
+
REP=$W/reports; mkdir -p "$REP" "$W/metrics"
|
| 16 |
+
LOG=$W/kld2048.log
|
| 17 |
+
exec > >(tee -a "$LOG") 2>&1
|
| 18 |
+
echo "==================== [$(date)] AMALIA KLD ctx2048 START ===================="
|
| 19 |
+
ppl_of(){ grep "Final estimate: PPL" "$1" | sed -E 's/.*PPL = ([0-9.]+).*/\1/' | head -1; }
|
| 20 |
+
|
| 21 |
+
bash $GL acquire agentic-install high "AMALIA KLD ctx2048" 2>&1 | tail -1
|
| 22 |
+
|
| 23 |
+
# (coherence check removed — llama-cli hangs on stdin; Q8's tiny KLD already proves F16 is sane)
|
| 24 |
+
|
| 25 |
+
# 1. F16 base at ctx 2048
|
| 26 |
+
echo "--- [1] F16 base (ctx 2048) ---"
|
| 27 |
+
"$PPX" -m "$F16" -f "$CORPUS" -c 2048 --kl-divergence-base "$KBASE" -ngl 99 > "$REP/base2048.txt" 2>&1 || { echo "base FAIL"; bash $GL release agentic-install; exit 1; }
|
| 28 |
+
F16PPL=$(ppl_of "$REP/base2048.txt")
|
| 29 |
+
echo " F16 PPL(ctx2048) = ${F16PPL}"
|
| 30 |
+
echo "name,size_gb,ppl,ppl_delta_vs_f16,kld_mean,kld_p95,kld_max,top1_match_pct" > "$CSV"
|
| 31 |
+
echo "F16,17.05,${F16PPL},0.0,0.0,0.0,0.0,100.0" >> "$CSV"
|
| 32 |
+
|
| 33 |
+
# 2. Each quant vs base at ctx 2048
|
| 34 |
+
for Q in Q8_0 Q6_K Q5_K_M Q4_K_M IQ4_XS Q3_K_M; do
|
| 35 |
+
M=$W/amalia-9b-${Q}.gguf; [ -f "$M" ] || { echo " $Q missing"; continue; }
|
| 36 |
+
R=$REP/kld2048-${Q}.txt
|
| 37 |
+
echo "--- [2] $Q (ctx 2048) ---"
|
| 38 |
+
"$PPX" -m "$M" -f "$CORPUS" -c 2048 --kl-divergence-base "$KBASE" --kl-divergence -ngl 99 > "$R" 2>&1
|
| 39 |
+
PPL=$(grep "Mean PPL(Q) " "$R"|grep -oE "[0-9]+\.[0-9]+"|head -1)
|
| 40 |
+
DELTA=$(grep "Mean PPL(Q)-PPL(base)" "$R"|grep -oE "\-?[0-9]+\.[0-9]+"|head -1)
|
| 41 |
+
KM=$(grep "Mean KLD:" "$R"|grep -oE "[0-9]+\.[0-9]+"|head -1)
|
| 42 |
+
KP95=$(grep "95.0% KLD:" "$R"|sed 's/.*KLD://'|grep -oE "[0-9]+\.[0-9]+"|head -1)
|
| 43 |
+
KMAX=$(grep "Maximum KLD:" "$R"|grep -oE "[0-9]+\.[0-9]+"|head -1)
|
| 44 |
+
T1=$(grep "Same top p:" "$R"|grep -oE "[0-9]+\.[0-9]+"|head -1)
|
| 45 |
+
SZ=$(du -m "$M"|awk '{printf "%.2f",$1/1024}')
|
| 46 |
+
echo " $Q: ${SZ}GB ppl=${PPL} dppl=${DELTA} kld=${KM} p95=${KP95} top1=${T1}%"
|
| 47 |
+
echo "${Q},${SZ},${PPL},${DELTA},${KM},${KP95},${KMAX},${T1}" >> "$CSV"
|
| 48 |
+
done
|
| 49 |
+
bash $GL release agentic-install 2>&1 | tail -1
|
| 50 |
+
echo "==================== [$(date)] DONE ===================="
|
| 51 |
+
cat "$CSV"
|
| 52 |
+
echo "KLD2048_COMPLETE"
|