-
training: 200 steps final run, lr=1e-5
-
training: add early-stop callback + lr=1e-5 for next run
-
training: add early-stop callback + lr=1e-5 for next run
-
Sync README.md
-
Sync pyproject.toml
-
Sync physix/ to merged tree (post train/ merge, pre 4ep/500step run)
-
add --system-ids flag; add job_train_single.py for damped_spring
-
Add checkpoint helpers + SFT hub push + GRPO resume
-
Add _render_training_curves + matplotlib dep + plot deliverables in README
-
7B run: anti-hack reward set + 7B profile
-
anti-hack rewards: gate simplicity on match>=.10, format on sim success, +match_dense +correctness_bonus, drop progress
-
drop redundant 'progress' reward (==match in single-turn); add lineage to W&B
-
loop.py: prominent W&B checkpoint logging (summary keys + table)
-
loop.py: PEFT-based adapter resume (overrides stale base path)
-
resume: --lora-adapter-repo support in loop.py
-
Initial physix-live source for HF Jobs training
-
initial commit
