Upload README.md with huggingface_hub

README.md

@@ -11,42 +11,139 @@ tags:
 - rewriting
 - style-transfer
 - unslop
-pipeline_tag: text-generation
+- text-generation
 ---
 
 # qwen3.5-0.8b-unslop-good-lora-v1
 
-Unslop rewrite adapter focused on reducing hype/corporate phrasing while preserving meaning.
+Smallest and cheapest lane in the Unslop family.
+
+This is a Qwen 3.5 0.8B fine-tune for rewrite-style cleanup: take AI-sounding prose, rewrite it into cleaner and more natural text, and keep the meaning intact.
+
+If you want the best quality in this family, start with 4B. If you want the lightest pilot or the lowest-cost baseline, this is the one.
+
+## Quick links
+
+- Hub repo: [Oysiyl/qwen3.5-0.8b-unslop-good-lora-v1](https://huggingface.co/Oysiyl/qwen3.5-0.8b-unslop-good-lora-v1)
+- GGUF files: [gguf/](./tree/main/gguf)
+
+## Recommended downloads
+
+If you just want the model files, the GGUF folder now contains only the final quantized artifacts:
+
+- `gguf/q2_k_gguf/Qwen3.5-0.8B.Q2_K.gguf`
+- `gguf/q4_k_m_gguf/Qwen3.5-0.8B.Q4_K_M.gguf`
+- `gguf/q6_k_gguf/Qwen3.5-0.8B.Q6_K.gguf`
+- `gguf/q8_0_gguf/Qwen3.5-0.8B.Q8_0.gguf`
+
+| Format | Best for | Notes |
+| --- | --- | --- |
+| `q6_k` | Default local use | Best balance for the 0.8B lane |
+| `q4_k_m` | Low-VRAM use | Smaller and faster, with a quality drop |
+| `q8_0` | Highest quality | Largest file size, most faithful among the GGUFs |
+| `q2_k` | Tiny / fastest | Emergency fallback only |
+
+My practical recommendation: download `q6_k` first, then keep `q4_k_m` around if you need a smaller fallback.
+
+## What this model is for
+
+Use this model if you want:
+
+- a cheap rewrite baseline
+- a fast pilot before scaling to 2B / 4B
+- a lightweight deployment candidate for simple cleanup tasks
+- a comparison point for judging whether a larger model is worth the extra cost
+
+## What it is not for
+
+Be careful with this lane if you need:
+
+- strict factual preservation on long inputs
+- the strongest style fidelity in the family
+- production-grade rewriting with minimal drift
+
+In this family, 0.8B is the roughest member. It can do real rewrites, but it is clearly less stable than 2B and 4B.
+
+## How it was trained
 
-## Model summary
-- Repo: `Oysiyl/qwen3.5-0.8b-unslop-good-lora-v1`
 - Base model: `Qwen/Qwen3.5-0.8B`
-- Adapter type: LoRA
-- Pipeline: text generation / rewrite style transfer
-- Current downloads (snapshot): 215
-
-## Intended use
-- Rewrite AI-sounding drafts into cleaner, more natural prose.
-- Keep meaning and key facts intact.
-- Use as a post-processing layer for longform and social text cleanup.
-
-## Limitations
-- Can still over-rewrite some passages.
-- Not guaranteed to improve factual accuracy.
-- Should be human-reviewed for fidelity-sensitive outputs.
-
-## Evaluation notes
-This card records this model as part of the Unslop family with a common quality goal: preserve meaning, reduce hype, and avoid hallucinated additions.
-
-## Usage (PEFT)
-```python
-from transformers import AutoModelForCausalLM, AutoTokenizer
-from peft import PeftModel
-
-base = "Qwen/Qwen3.5-0.8B"
-adapter = "Oysiyl/qwen3.5-0.8b-unslop-good-lora-v1"
-
-tokenizer = AutoTokenizer.from_pretrained(base, trust_remote_code=True)
-base_model = AutoModelForCausalLM.from_pretrained(base, trust_remote_code=True, device_map="auto")
-model = PeftModel.from_pretrained(base_model, adapter)
-```
+- Training path: Transformers / TRL / PEFT fine-tuning on Hugging Face Jobs
+- Dataset: `N8Programs/unslop-good`
+- Rows used: 1000 (full training split)
+- Objective: direct rewrite / style cleanup
+
+## Training shape
+
+- hardware: A10G 24GB (`a10g-large`)
+- max_seq_length: 2048
+- num_train_epochs: 2
+- batch_size: 1
+- gradient_accumulation_steps: 1
+- learning_rate: 1e-4
+- scheduler: cosine
+- warmup_steps: 50
+- LoRA rank: 8
+- LoRA alpha: 20
+- LoRA dropout: 0.0
+- 4-bit loading
+- bf16 training
+
+## Deployment
+
+This run is deployment-backed and live on Modal.
+
+- Status: healthy
+- Base model: `Qwen/Qwen3.5-0.8B`
+- Output repo: `Oysiyl/qwen3.5-0.8b-unslop-good-lora-v1`
+- Mode: standard Unsloth / HF adapter loading
+- Endpoint style: rewrite API
+- Thinking mode: disabled
+
+Live endpoint generation defaults used by the app:
+
+- `enable_thinking = false`
+- `temperature = 0.7`
+- `top_p = 0.8`
+- `top_k = 20`
+- `min_p = 0.0`
+- `presence_penalty = 1.5`
+- `repetition_penalty = 1.0`
+
+## Evaluation snapshot
+
+Short eval summary only (full long-form input/output removed due poor quality).
+
+Short sanity sample:
+
+> Rewrite this short-form promo hook to sound less generic and more human.
+
+Full observed output:
+
+> The old ad said everything about being the best in town, but that's just a cliché. I'm not trying to be the king of the city. I'm just here to show you how we actually do things. If you want to see the real stuff, come by and let me show you what makes us special. Let's get started.
+
+Held-out fiction sample (input):
+
+> The wind rose all at once, as if the mountain itself had inhaled. She stood at the edge of the pass, cloak snapping like a torn banner behind her, staring down at the valley where the last lights were going out one by one. Each flame extinguished felt deliberate—an answer to a question she had been too afraid to ask. Somewhere below, bells began to ring, not in warning, but in mourning. They already knew. “Don’t,” he said, breathless, reaching for her arm. His fingers brushed fabric, then air. “If you cross that line, there’s no coming back.” She didn’t turn. The cold burned her lungs, sharpened her thoughts. “There was never a way back,” she replied softly. “Only a way forward you were willing to pretend didn’t exist.” Thunder rolled, low and ancient. For a moment, the sky fractured into white fire, and in that instant he saw her face—resolved, terrified, incandescent with purpose. Not the girl he had sworn to protect, but the woman the world had demanded she become. “You’ll die,” he said. She smiled then, finally facing him, eyes bright with unshed tears. “So will everyone,” she answered. “The difference is what we leave burning when we’re gone.” She stepped forward. The mountain exhaled.
+
+## Judgment
+
+Blunt judgment: this is the weakest lane in the family, but it is still a real model, not a toy. It can rewrite long prose in a coherent way, just not as reliably or as naturally as the larger lanes.
+
+Use it when you care more about cost and latency than final quality.
+
+## Family position
+
+- 0.8B: cheapest pilot, roughest output
+- 2B: better balance of fidelity and fluency
+- 4B: strongest default candidate in the small-model set
+- 9B: useful retrain, but still not the safest long-form rewrite choice versus 4B
+
+## Training loss vs progress
+
+See the normalized family comparison plot below.
+
+![Normalized training loss comparison: 9B vs 0.8B vs 2B vs 4B](./training_loss_vs_progress_comparison_9b_0_8b_vs_2b_vs_4b.svg)
+
+## Bottom line
+
+If you want the lightest Unslop lane, this is the one to grab. If you want higher rewrite quality, scale up to the larger lanes (9B and especially 30B-A3B for quality-first use).