---
license: gemma
datasets:
- GBaker/MedQA-USMLE-4-options-hf
base_model:
- google/gemma-3-12b-it
library_name: transformers
tags:
- biology
- medical
---
# Gemma-3-12B-GRPO trained with GRPO via LoRA

Due to limited available computational resources, we randomly sampled 500 data points from MedQA-USMLE using a methodology and conducted preliminary GRPO experiments with LoRA using the [Unsloth](https://github.com/unslothai/unsloth) framework. We are now releasing this as a preview version. More experiments and explorations are currently underway, and a technical report is in preparation. Thank you for your patience. We conduct the experiments on one RTX-A6000 Ada (48GB VRAM).

## Evaluation Results

The model is evaluated on four benchmark datasets: MMLU, MMLU-Pro, CMMU, GSM8K, GPQA. The experimental results are summarized in Table 1, with comprehensive analyses provided in the Detailed Results section.

<center><strong>Tab.1 Evaluation results.</strong></center>

| Dataset | Gemma-3-12b-it | Gemma3-12b-GRPO |
| :-----: | :------------: | :-------------: |
|  MMLU   |     65.51      |     70.13      |
|  MMLU-Pro   |     60.17      |     59.99      |
|  CMMLU   |     54.81     |     57.07     |
|  GSM8K   |     91.58     |     91.81     |
|  GPQA   |     34.98     |     34.23     |

## Requirements

```shell
pip install torch==2.6.0 torchaudio==2.6.0 torchvision==0.21.0 -i --index-url https://download.pytorch.org/whl/cu124
pip install transformer vllm bitsandbytes peft
pip install flash-attn --no-build-isolation
```

## Run with vLLM

You can use the following script to run with vLLM.

```shell
vllm serve qiuxi337/gemma-3-12b-it-grpo \
     --gpu-memory-utilization 0.85 \
     --max-model-len 4096 \
     --served-model-name gemma3-12b-grpo \
     --api-key your_api_key
```

## Detail Results

### MMLU

![mmlu](asserts/mmlu.png)

**Fig.1 The results on the MMLU benchmark.**

![MMLU_Humanities](asserts/mmlu_humanities.png)

**Fig.2  The results on the MMLU-Humanities**

![MMLU_Social_Science](asserts/mmlu_social_science.png)

**Fig.3  The results on the MMLU-Social Science**

![MMLU_STEM](asserts/mmlu_stem.png)

**Fig.4  The results on the MMLU-STEM**

![MMLU_others](asserts/mmlu_other.png)

**Fig.5  The results on the MMLU-Other**

### MMLU-Pro

![MMLU_Pro](asserts/mmlu_pro.png)

**Fig.6  The results on the MMLU-Pro**

### CMMLU

![cmmlu](asserts/cmmlu.png)

**Fig.7  The results on the CMMLU benchmark.**

![CMMLU_Humanities](asserts/cmmlu_humanities.png)

**Fig.8  The results on the CMMLU-Humanities**

![CMMLU_Social_Science](asserts/cmmlu_social_science.png)

**Fig.9  The results on the CMMLU-Social Science**

![CMMLU_STEM](asserts/cmmlu_stem.png)

**Fig.10  The results on the CMMLU-STEM**

![CMMLU_others](asserts/cmmlu_other.png)

**Fig.11  The results on the CMMLU-Other**

![CMMLU_China_Specific](asserts/cmmlu_china_specific.png)

**Fig.12  The results on the CMMLU-China Specific**

## Acknowledge

[Gemma-3-12b-it](https://huggingface.co/google/gemma-3-12b-it)

[Unlsoth](https://github.com/unslothai/unsloth)

## Citation

```ini
@software{Qiu_Open-Medical-R1,
author = {Qiu, Zhongxi and Zhang, Zhang and Hu, Yan and Li, Heng and Liu, Jiang},
license = {MIT},
title = {{Open-Medical-R1}},
url = {https://github.com/Qsingle/open-medical-r1},
version = {0.1}
}
```