sync training files

fix: OOM — NUM_GENERATIONS 32→16, max_completion_length 300→200, expandable_segments

fix: add 3 missing hard tasks to _TASK_DIFFICULTY (322 prompts)

Fix: kl_coeff -> beta (correct TRL GRPOConfig param name)

Auto-detect username from token for adapter + run folder upload

Graceful stop: save weights on /app/stop_requested flag

Fix hard_currency_conversion task ID in TRAIN_TASKS and EVAL_TASKS

Add oversight eval + bump seeds medium×8 hard/long×20

Bump seeds: medium×8, hard/long×20 (322 prompts total)

Full metrics in live JSON: format/diff/ep_len history

Add loss tracking callback + reward/loss PNG savers

Run 3: temp=0.7, kl=0.1, format±0.15, no curriculum, 20 tasks, G=32

Fix bf16 crash + 17 tasks / 160 prompts dataset

Upload train.py with huggingface_hub

Upload train.py with huggingface_hub

Upload train.py with huggingface_hub

Upload train.py with huggingface_hub

Upload train.py with huggingface_hub

Upload train.py with huggingface_hub

Upload train.py with huggingface_hub