Views
No views yet
Qwen/Qwen3-4B with GRPO on the tooluse split.val-aux/tooluse/reward/mean@16 from 10-step validation logs.| best mean@16 | best step | final mean@16 | final step |
|---|---|---|---|
| 61.31% | 80 | 60.29% | 100 |
| step | mean@16 |
|---|---|
| 10 | 58.00% |
| 20 | 60.11% |
| 30 | 61.21% |
| 40 | 60.11% |
| 50 | 60.94% |
| 60 | 59.28% |
| 70 | 58.64% |
| 80 | 61.31% |
| 90 | 57.72% |
| 100 | 60.29% |
metrics.json: parsed validation summaryeval_mean16.csv: step-level validation curve dataeval_mean16.png: validation curve plotglobal_step_100/actor checkpoint. If best step is earlier than 100, the best validation point is reported for tracking, but the corresponding actor weights may not be retained locally./mnt/mole/SDPO/L2T/checkpoints/datasets/tooluse/qwen3gen-tooluse-GRPO-Qwen-Qwen3-4B-mbs8-train64-rollout8-lr1e-6-vllm0.8run-20260630_031808-osu8wla2global_step_100/actor converted from VERL FSDP shards to Hugging Face format.