---
license: apache-2.0
library_name: transformers
pipeline_tag: text-generation
tags:
- qwen3
- reinforcement-learning
- rlsd
---

# qwen3-4b-material-rlsd-ema005

Fine-tuned from `Qwen/Qwen3-4B` with RLSD (EMA 0.05) on the `material` 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 |
|---:|---:|---:|---:|
| 77.19% | 60 | 76.26% | 100 |

![Validation mean@16](eval_mean16.png)

| step | mean@16 |
|---:|---:|
| 10 | 63.76% |
| 20 | 74.07% |
| 30 | 76.60% |
| 40 | 75.86% |
| 50 | 76.73% |
| 60 | 77.19% |
| 70 | 72.94% |
| 80 | 73.01% |
| 90 | 74.00% |
| 100 | 76.26% |

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/material/qwen3gen-material-RLSD-Qwen-Qwen3-4B-mbs8-decay0-ema0.05-train64-rollout8-lr1e-6-vllm0.8`

W&B run: `run-20260630_014254-tpv9xlh5`

This upload uses `global_step_100/actor` converted from VERL FSDP shards to Hugging Face format.