Squash-merge of run/co_pretraining_2026_04_13 into main.
Replaces the legacy backbone (4,278-token coordinate vocab, 256 context,
prepend_outcome=True, 100K steps) with the v1.0.0 backbone:
- 1,980-token searchless_chess action vocabulary (1,968 reachable actions
+ 1 PAD + 11 outcome tokens)
- 512-token context window
- prepend_outcome=False (no outcome conditioning)
- 200K training steps at batch size 256 (best checkpoint at step 195K)
- 66909440 parameters
Published model.safetensors is the best 5K-cadence checkpoint by val
loss (step 195,000). All intermediate 5K checkpoints are preserved under
checkpoints/ for training-dynamics analysis.
The legacy checkpoint is preserved in the git history of this repo and
is also available at thomas-schweich/pawn-large-legacy.
