Gala: an Apple-Silicon-first model, trained on one
junafinity
• • 1image imagewidth (px) 1.28k 1.53k |
|---|
Figures for the Gala article (see junafinity/Gala-598M-MLX). All plots are generated directly from the training/eval logs shipped in the model repo.
| file | shows |
|---|---|
fig_10m.png |
10.5M-token stream: decode speed and next-2k loss flat at every depth |
fig_seeds.png |
3-seed comparison, Gala vs Transformer, matched 120M tokens |
fig_scale.png |
120M→300M scale study: the curves diverge |
fig_passkey.png |
recall boundary: 100% in-window, chance beyond |
fig_decode.png |
decode speed vs context, constant state vs growing KV |
fig_nan.png |
the chunked-delta-rule inverse instability trace |