fprm-flagship-L4

FPRM flagship L4 — two arms (Sudoku-Extreme)

Raw research checkpoints for two flagship runs. Accuracy numbers are not included here on purpose: they are re-measured with the frozen eval pipeline.

IMPORTANT: use each run's own trm.py

The two arms were trained with different trm.py revisions (differing md5). fl-carrylast-L4-s1's behavioural difference lives in its code, not in its config or parameters — its state dict has no loop_attn.* params at any step (verified at 5000 / 45570 / 78120, so no mid-run architecture change). Each directory ships the exact trm.py, losses.py and all_config.yaml it was trained with. Do not cross-load.

Shared config

data/sudoku-extreme-1k-aug-1000, hidden 512, L_layers=4, H_cycles=3, L_cycles=6, halt_max_steps=16, rope, bf16, stablemax_cross_entropy, q_loss_coeff=0.5. Full config in each all_config.yaml.

Files per run

• step_<N> — model state dict only (109 MB)
• step_<N>_train_state.pt — model + optimizer, for resuming (273 MB)
• all_config.yaml, trm.py, losses.py

Training logs (logs/)

Both runs reached 78,125 steps and ended exit=0 / DONE. No NaN and no OOM in either log. Both were restarted and resumed from checkpoints mid-training. The b01 log contains 6 tracebacks, all infrastructure-level, none a training failure: 4x ModuleNotFoundError: coolname (launch-env, fast-failed and relaunched) and a disk-full (Errno 28) while writing step_65100 on 2026-07-17, after which the run resumed and completed.

Last logged train metrics (not eval):