WORKER · QWEN2.5-1.5B · GRPO

Watch the learning signal unfold.

A timed replay of 216 measurements from the original training run. No model is training now and no values are generated by this page.

STEP000 / 216optimizer update

REWARD—group mean

LOSS—optimization signal

KL—policy divergence

MEAN LENGTH—output tokens

Speed

Total reward

Higher is better

Optimizer output

Distance from reference

Mean generated tokens