# Training Curves

PNGs in this directory are auto-generated by
`physix.training.loop._render_training_curves` at end of every GRPO run, then
mirrored from the HF model repo via `train/sync-plots.sh`.

Files:

- `loss.png`              — GRPO surrogate loss over training steps.
- `reward.png`            — Mean reward (with ±1σ band) over training steps.
- `reward_components.png` — Per-component reward (`match`, `match_dense`,
                             `correctness`, `simplicity`, `format`).

To regenerate locally after a job:

    ./train/sync-plots.sh Pratyush-01/physix-3b-rl