figures for the Gala article
Browse files- README.md +18 -0
- fig_10m.png +3 -0
- fig_decode.png +3 -0
- fig_nan.png +3 -0
- fig_passkey.png +3 -0
- fig_scale.png +3 -0
- fig_seeds.png +3 -0
README.md
ADDED
|
@@ -0,0 +1,18 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
pretty_name: Gala article assets
|
| 4 |
+
---
|
| 5 |
+
|
| 6 |
+
# gala-article-assets
|
| 7 |
+
|
| 8 |
+
Figures for the Gala article (see [junafinity/Gala-598M-MLX](https://huggingface.co/junafinity/Gala-598M-MLX)).
|
| 9 |
+
All plots are generated directly from the training/eval logs shipped in the model repo.
|
| 10 |
+
|
| 11 |
+
| file | shows |
|
| 12 |
+
|---|---|
|
| 13 |
+
| `fig_10m.png` | 10.5M-token stream: decode speed and next-2k loss flat at every depth |
|
| 14 |
+
| `fig_seeds.png` | 3-seed comparison, Gala vs Transformer, matched 120M tokens |
|
| 15 |
+
| `fig_scale.png` | 120M→300M scale study: the curves diverge |
|
| 16 |
+
| `fig_passkey.png` | recall boundary: 100% in-window, chance beyond |
|
| 17 |
+
| `fig_decode.png` | decode speed vs context, constant state vs growing KV |
|
| 18 |
+
| `fig_nan.png` | the chunked-delta-rule inverse instability trace |
|
fig_10m.png
ADDED
|
Git LFS Details
|
fig_decode.png
ADDED
|
Git LFS Details
|
fig_nan.png
ADDED
|
Git LFS Details
|
fig_passkey.png
ADDED
|
Git LFS Details
|
fig_scale.png
ADDED
|
Git LFS Details
|
fig_seeds.png
ADDED
|
Git LFS Details
|