gemma-3-12b-it-grpo

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 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>

Requirements

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.

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

Fig.1 The results on the MMLU benchmark.

Fig.2 The results on the MMLU-Humanities

Fig.3 The results on the MMLU-Social Science

Fig.4 The results on the MMLU-STEM

Fig.5 The results on the MMLU-Other

MMLU-Pro

Fig.6 The results on the MMLU-Pro

CMMLU

Fig.7 The results on the CMMLU benchmark.

Fig.8 The results on the CMMLU-Humanities

Fig.9 The results on the CMMLU-Social Science

Fig.10 The results on the CMMLU-STEM

Fig.11 The results on the CMMLU-Other

Fig.12 The results on the CMMLU-China Specific

Acknowledge

Gemma-3-12b-it

Unlsoth

Citation

@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}
}