---
license: apache-2.0
base_model: Qwen/Qwen3-4B
library_name: transformers
pipeline_tag: text-generation
tags:
- qwen3
- chemistry
- reinforcement-learning
- verl
- sciknoweval
- sdpo
- self-distillation
---

# Qwen3-4B Chemistry SDPO

This repository contains Chemistry fine-tuned Qwen3-4B checkpoints from the local SciKnowEval-style generalization setup.

- Root checkpoint: final `global_step_100` merged to Hugging Face safetensors.
- `best_avg16/`: checkpoint with the highest validation `avg@16` during training, merged to Hugging Face safetensors.

## Checkpoints

| Checkpoint | Source step | Validation avg@16 | best@16 | maj@16 |
|---|---:|---:|---:|---:|
| Root final | 100 | 0.720536 | 0.731567 | 0.721367 |
| `best_avg16/` | 20 | 0.766369 | 0.822871 | 0.779486 |

## Training Run

`qwen3gen-chemistry-SDPO-Qwen-Qwen3-4B-mbs32-ema0.05-train256-rollout8-lr1e-5-vllm0.8`

W&B run: `https://wandb.ai/seongryongjung-chung-ang-university/qwen3-generalization-batch256/runs/1qazekk3`

## Base Model

- Base model: `Qwen/Qwen3-4B`
- Fine-tuning type: full-parameter FSDP RL training
- Dataset: `datasets/sciknoweval/chemistry`
- Train split: 1,890 examples
- Validation split: 210 examples

## Method

- Method: SDPO
- Config: `sdpo`
- Policy loss mode: `sdpo`
- Reward: local SciKnowEval multiple-choice reward checker
- Rollout correction: token-level importance sampling, threshold 2.0

## Hyperparameters

| Field | Value |
|---|---:|
| Base model | `Qwen/Qwen3-4B` |
| Training steps | 100 |
| Train batch size | 256 |
| Rollouts per prompt | 8 |
| Generations per step | 2048 |
| PPO mini batch size | 32 |
| Learning rate | `1e-5` |
| LR warmup steps | 10 |
| Weight decay | 0.01 |
| Grad clip | 1.0 |
| Max prompt length | 2048 |
| Max response length | 8192 |
| Max model length | 10240 |
| Train temperature | 1.0 |
| Train top_p | 1.0 |
| Validation generations | 16 |
| Validation temperature | 0.6 |
| Validation top_p | 0.95 |
| vLLM GPU memory utilization | 0.8 |
| GPUs | 8 x NVIDIA H200 |
| Save frequency | every 10 steps |
| Validation frequency | every 10 steps |
| Distillation top-k | 100 |
| SDPO alpha | 0.5 |
| Teacher update rate | 0.05 |
| Distillation IS clip | 2.0 |
| Max reprompt length | 10240 |

## Metrics

![Training score](training_score.png)

CSV files:

- [`training_score.csv`](training_score.csv)
- [`validation_metrics.csv`](validation_metrics.csv)

| Metric | Value |
|---|---:|
| Final training step | 100 |
| Final `critic/score/mean` | 0.845215 |
| Final `critic/rewards/mean` | 0.845215 |
| Final validation `avg@16` | 0.720536 |
| Peak validation `avg@16` | 0.766369 |
| Peak validation step | 20 |

## Loading

Root final checkpoint:

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("SeongryongJung/Qwen3-4B-Chemistry-SDPO")
tokenizer = AutoTokenizer.from_pretrained("SeongryongJung/Qwen3-4B-Chemistry-SDPO")
```

Best avg@16 checkpoint:

```python
model = AutoModelForCausalLM.from_pretrained("SeongryongJung/Qwen3-4B-Chemistry-SDPO", subfolder="best_avg16")
tokenizer = AutoTokenizer.from_pretrained("SeongryongJung/Qwen3-4B-Chemistry-SDPO", subfolder="best_avg16")
```

## Intended Use

This model is intended for research on RL fine-tuning and self-distillation behavior on science/generalization tasks. It has not been broadly safety evaluated for production use.

## Limitations

The reported scores are training-time and validation-time metrics from the local experimental setup. They should not be interpreted as broad benchmark results without independent evaluation.