---
license: apache-2.0
base_model: google/medgemma-4b-it
tags:
  - medical
  - safety
  - grpo
  - dipg
  - safeclaw
library_name: transformers
pipeline_tag: text-generation
---

# MedGemma 4B - DIPG Safety v2 (GRPO-Trained)

Fine-tuned [MedGemma 4B](https://huggingface.co/google/medgemma-4b-it) using 
**GRPO (Group Relative Policy Optimization)** on the DIPG Safety Gym benchmark.

## Training Details
- **Base Model**: MedGemma 4B IT
- **Method**: GRPO with LoRA (rank=64, alpha=64)
- **Training Steps**: 100
- **Focus**: Medical safety — hallucination reduction, evidence-grounded responses

## Usage

```python
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("surfiniaburger/medgemma-4b-dipg-safety-v2")
model = AutoModelForCausalLM.from_pretrained(
    "surfiniaburger/medgemma-4b-dipg-safety-v2",
    torch_dtype="bfloat16",
    device_map="auto"
)
```

## SafeClaw Project
Part of the [DIPG Safety Gym](https://github.com/surfiniaburger/med-safety-gym-v2) ecosystem.