Add periodic safe eval during training#329

Closed

eugenevinitsky wants to merge 43 commits into3.0_betafrom

safe-eval-during-training

Commits on Mar 8, 2026

Add periodic safe eval during training with reward conditioning

Eugene Vinitsky
authored and
eugenevinitsky
committed
Move bin file cleanup from render_videos to caller

eugenevinitsky
and
claude
committed

Commits on Mar 9, 2026

Fix eval subprocess: pass device and handle negative arg values

eugenevinitsky
and
claude
committed
Apply ruff formatting to utils.py

eugenevinitsky
and
claude
committed
Decouple render and safe eval from checkpoint interval

eugenevinitsky
and
claude
committed
Simplify safe_eval config: remove redundant prefixes

eugenevinitsky
and
claude
committed
Replace SAFE_EVAL_REWARD_BOUNDS list with dynamic config iteration

eugenevinitsky
and
claude
committed
Discover reward bounds from env config instead of hardcoding

eugenevinitsky
and
claude
committed
Fix wosac eval: use correct CLI flag and show more stderr on failure

eugenevinitsky
and
claude
committed
Pass episode_length=91 to wosac eval subprocess to match ground truth logs

eugenevinitsky
and
claude
committed
Add eval_async option to run eval subprocesses in background threads

eugenevinitsky
and
claude
committed
Fix ruff formatting for eval_async call sites

eugenevinitsky
and
claude
committed

Commits on Mar 13, 2026

Commits on Mar 14, 2026

Commits on Mar 15, 2026

Commits on Mar 16, 2026