51 lines
1.3 KiB
Markdown
51 lines
1.3 KiB
Markdown
---
|
|
license: apache-2.0
|
|
library_name: transformers
|
|
pipeline_tag: text-generation
|
|
tags:
|
|
- qwen3
|
|
- reinforcement-learning
|
|
- grpo
|
|
---
|
|
|
|
# qwen3-4b-physics-grpo
|
|
|
|
Fine-tuned from `Qwen/Qwen3-4B` with GRPO on the `physics` split.
|
|
|
|
## Validation Performance
|
|
|
|
Metric: `val-aux/sciknoweval/reward/mean@16` from 10-step validation logs.
|
|
|
|
| best mean@16 | best step | final mean@16 | final step |
|
|
|---:|---:|---:|---:|
|
|
| 62.27% | 20 | 53.05% | 100 |
|
|
|
|
|
|
|
|
| step | mean@16 |
|
|
|---:|---:|
|
|
| 10 | 60.00% |
|
|
| 20 | 62.27% |
|
|
| 30 | 46.88% |
|
|
| 40 | 52.34% |
|
|
| 50 | 58.91% |
|
|
| 60 | 61.80% |
|
|
| 70 | 52.34% |
|
|
| 80 | 56.17% |
|
|
| 90 | 53.44% |
|
|
| 100 | 53.05% |
|
|
|
|
Files included with this repo:
|
|
- `metrics.json`: parsed validation summary
|
|
- `eval_mean16.csv`: step-level validation curve data
|
|
- `eval_mean16.png`: validation curve plot
|
|
|
|
Important: the uploaded weights are the final `global_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.
|
|
|
|
Checkpoint source:
|
|
`/mnt/mole/SDPO/L2T/checkpoints/datasets/sciknoweval/physics/qwen3gen-physics-GRPO-Qwen-Qwen3-4B-mbs8-train64-rollout8-lr1e-6-vllm0.8`
|
|
|
|
W&B run: `run-20260629_174849-hz5zwriz`
|
|
|
|
This upload uses `global_step_100/actor` converted from VERL FSDP shards to Hugging Face format.
|