---
license: apache-2.0
base_model: Qwen/Qwen3-4B
library_name: transformers
tags:
- qwen3
- reinforcement-learning
- trust-region
- sdpo
- srpo
---

# Qwen3-4B-Chemistry-SRPO-TR

This repository contains the Qwen3-4B `Chemistry` `SRPO-TR` run trained with batch size 32.

The repository root contains the **best validation checkpoint** (`global_step_100`), selected by validation `mean@16`.
The best validation checkpoint is also the final checkpoint, so the repository root is both best and last.

## Performance

| Dataset | Method | Base model | Train batch size | Best val mean@16 | Best checkpoint | Final val mean@16 | Final checkpoint |
|---|---|---|---:|---:|---:|---:|---:|
| Chemistry | SRPO-TR | Qwen3-4B | 32 | 76.61% | 100 | 76.61% | 100 |

![Training and validation scores](results/training_score.png)

Raw result files:

- `results/validation_mean16.csv`
- `results/training_scores.csv`
- `artifacts/config.yaml`
- `artifacts/wandb-summary.json`

## Training Hyperparameters

| Hyperparameter | Value |
|---|---:|
| Base model | `Qwen/Qwen3-4B` |
| Method | `SRPO-TR` |
| Train batch size | `32` |
| Train max samples | `3200` |
| Total training steps | `100` |
| Save frequency | `10` |
| Test frequency | `10` |
| Teacher regularization | `trust-region` |
| Teacher update rate | `0.1` |
| Rollout samples per prompt | `8` |
| Validation samples per prompt | `16` |
| Learning rate | `5e-6` |
| vLLM GPU memory utilization | `0.8` |

## Usage

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

repo_id = "SeongryongJung/Qwen3-4B-Chemistry-SRPO-TR"
tokenizer = AutoTokenizer.from_pretrained(repo_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    repo_id,
    torch_dtype="auto",
    device_map="auto",
    trust_remote_code=True,
)
```