Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,48 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
tags:
|
| 3 |
+
- pytorch
|
| 4 |
+
- language-model
|
| 5 |
+
- gpt
|
| 6 |
+
license: mit
|
| 7 |
+
---
|
| 8 |
+
|
| 9 |
+
# vanilla-10b
|
| 10 |
+
|
| 11 |
+
Vanilla GPT baseline trained to compare against [aemack-org/cayley-10b](https://huggingface.co/aemack-org/cayley-10b).
|
| 12 |
+
|
| 13 |
+
## Architecture
|
| 14 |
+
|
| 15 |
+
| Parameter | Value |
|
| 16 |
+
|-----------|-------|
|
| 17 |
+
| n_layer | 12 |
|
| 18 |
+
| n_head | 8 |
|
| 19 |
+
| n_embd | 1024 |
|
| 20 |
+
| block_size | 1024 |
|
| 21 |
+
| vocab_size | 50304 |
|
| 22 |
+
| bias | False |
|
| 23 |
+
| norm | RMSNorm (affine) |
|
| 24 |
+
| MLP | GELU, 4x expansion |
|
| 25 |
+
| tokenizer | GPT-2 (tiktoken) |
|
| 26 |
+
| dtype | bfloat16 |
|
| 27 |
+
| sparsity | none (vanilla) |
|
| 28 |
+
|
| 29 |
+
## Training
|
| 30 |
+
|
| 31 |
+
| Parameter | Value |
|
| 32 |
+
|-----------|-------|
|
| 33 |
+
| optimizer | Muon (hidden 2D) + AdamW (embeddings) |
|
| 34 |
+
| muon_lr | 0.006 |
|
| 35 |
+
| adamw_lr | 0.006 |
|
| 36 |
+
| lr_schedule | linear_warmdown (warmdown_frac=0.2) |
|
| 37 |
+
| batch_size | 80 |
|
| 38 |
+
| seq_len | 1024 |
|
| 39 |
+
| max_iters | 16000 |
|
| 40 |
+
| tokens seen | ~1.3B |
|
| 41 |
+
| dataset | FineWeb-Edu-10B |
|
| 42 |
+
| best_val_loss | 6.2834 |
|
| 43 |
+
|
| 44 |
+
## Purpose
|
| 45 |
+
|
| 46 |
+
Interpretability comparison baseline. Trained with identical hyperparameters to
|
| 47 |
+
`cayley-10b` but without the CayleySAE bottleneck at `mlp_in`. Enables direct
|
| 48 |
+
comparison of residual stream representations.
|