3dttt-clm-ft1

Fine-tuned choice head for CLM at 4×4×4 (4-in-a-row) 3D tic-tac-toe.

This is the ft1 round (2026-10-01): the CLM_v0.1-8B encoder was frozen and only the state/action head was trained on teacher-oracle-labeled self-play.

This repository hosts only the head (best_head.pt). It is not a standalone model — it must be served on top of the CLM base with the clm-serve tooling from the training workspace.

Files

file	size	md5	sha256
best_head.pt	75,557,470 B	911b68ca503fc6f1a326233b7107fac0	c37c2534b80d4ab5aac6be0ff9323e3b6430f784770b3ef4c61d49a99b648934

Training

• Data: oracle-labeled 3D tic-tac-toe self-play (3dttt-v1): 44,762 unique positions (99,648 with block/win/threat boosting), 179,366 questions; 5,017 held-out test positions.
• Config: frozen CLM encoder, head-only training; best epoch 20 (val acc 0.9412); ~51 min wall time on one GPU pod.
• Selection: best held-out validation accuracy. The file was recovered from the training pod after the run; the hashes above pin its provenance.

Metrics (held-out test)

model	test acc	answer	threat_check	soft CE
init (zero-shot)	0.2925	0.2511	0.3339	1.5312
majority class	0.2973	—	—	—
ft1 best_head	0.8917	0.7833	1.0000	0.3096

Targeted nuke double-threat probe: nuke probability 0.993.

Head-to-head vs JEV (valid re-run, 2026-10-04)

40 games, 10 seeds, fail-fast runner, 0 invalid moves:

	Games	3dttt-clm-ft1	JEV
Overall	40	24W-13L-3D (win-equiv 64%, 95% CI 48–77%)	13W-24L-3D
as X (first)	20	20W-0L-0D	0W-20L
as O (second)	20	4W-13L-3D	13W-4L-3D

An earlier ft1-vs-JEV series (2026-10-01) was invalidated (stale server environment silently fell back to random moves) and must not be cited.

Usage

wget https://huggingface.co/DeKodez/3dttt-clm-ft1/resolve/main/best_head.pt
# serve on top of CLM_v0.1-8B (training workspace tooling):
clm-serve --ckpt best_head.pt