Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,31 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
language:
|
| 3 |
+
- en
|
| 4 |
+
license: apache-2.0
|
| 5 |
+
tags:
|
| 6 |
+
- glove
|
| 7 |
+
- lora
|
| 8 |
+
- distillation
|
| 9 |
+
- bpe
|
| 10 |
+
- cl100k_base
|
| 11 |
+
- ffn
|
| 12 |
+
base_model: jsanzolac/bpe_glove_512
|
| 13 |
+
datasets:
|
| 14 |
+
- jsanzolac/qwen3_emb_512
|
| 15 |
+
- jsanzolac/qwen3_emb_512_packed
|
| 16 |
+
---
|
| 17 |
+
|
| 18 |
+
# bpe_glove_512_lora_v1_ffn
|
| 19 |
+
|
| 20 |
+
Warm-start from `jsanzolac/bpe_glove_512_lora_v1/rank_512` plus a per-token FFN inserted
|
| 21 |
+
between the GloVe-attention output and the alpha-pool collapse.
|
| 22 |
+
|
| 23 |
+
**Trainable:** `A`, `B`, FFN. **Frozen:** `E`, teacher.
|
| 24 |
+
**Loss:** `λ_c·InfoNCE + λ_D·‖ρ_T − ρ_S‖²_F` with `λ_c=1.0`, `λ_D=0.1`.
|
| 25 |
+
Density is computed on the **post-FFN** per-token states; InfoNCE is on the alpha-pooled sentence vector.
|
| 26 |
+
|
| 27 |
+
Files:
|
| 28 |
+
- `rank_512/checkpoint_final.pt` — A + B + FFN state dict (E is non-persistent; re-inject from `jsanzolac/bpe_glove_512/vectors.txt`).
|
| 29 |
+
- `rank_512/config.json` — full hyperparameters.
|
| 30 |
+
- `rank_512/vectors_drifted.txt` — `E + B(A(·))` per vocab row, GloVe text format. Note: this captures only the static drifted embedding lookup, **not** the FFN's effect (which is contextual). To use the model end-to-end, instantiate `DriftingGloVeStudentFFN` and run forward.
|
| 31 |
+
- `rank_512/train_log.jsonl` — per-step metrics.
|