chess-sim: RL pilot (ReinFlow-style PPO on SmolVLA), stopped after round 7: checks, logs, round-4 checkpoint
e534973 verified | # Held-out test (fixed before the pilot): the 200 reserved fresh scenes (seed 9100003), anti-aliasing off, | |
| # eval_policy.py --fast (checked identical to the normal evaluation). | |
| # primary: RL checkpoint with the best validation success vs the baseline, both with the action | |
| # expert in float32 (the precision RL trains and runs it in) | |
| # secondary: the baseline as always evaluated (expert rounded to bfloat16 on load) | |
| cd /root/play_machanize | |
| export MUJOCO_GL=egl OPENBLAS_NUM_THREADS=1 PHASE2_PROFILE=baseline SIM_OFFSAMPLES=0 | |
| R=sim/reports/rl/heldout; mkdir -p $R | |
| RL=${RL_CKPT:-outputs/rl_pilot/best} | |
| E=".venv/bin/python sim/eval_policy.py --fast --episodes 200 --workers 4 --seed 9100003 --video-episodes 2" | |
| $E --expert-fp32 --policy models/baseline --out $R/baseline_fp32 > logs/heldout_baseline_fp32.log 2>&1 & | |
| $E --expert-fp32 --policy $RL --out $R/rl > logs/heldout_rl.log 2>&1 | |
| wait | |
| $E --policy models/baseline --out $R/baseline_native > logs/heldout_baseline_native.log 2>&1 | |
| .venv/bin/python sim/paired_compare.py baseline_fp32=$R/baseline_fp32 rl=$R/rl --out $R/paired.md | |
| .venv/bin/python sim/paired_compare.py baseline_native=$R/baseline_native rl=$R/rl --out $R/paired_vs_native.md | |
| echo HELDOUT_DONE | |