README.md

---
tags:
- safety
- sft
- gemma
license: apache-2.0
datasets:
- nvidia/Aegis-AI-Content-Safety-Dataset-2.0
language:
- en
---

# google/gemma-7b-it Safety SFT

This model is fine-tuned from `google/gemma-7b-it` using the Nvidia Aegis AI Content Safety Dataset 2.0.

## Training Details
- **Base Model**: google/gemma-7b-it
- **Dataset**: nvidia/Aegis-AI-Content-Safety-Dataset-2.0
- **Training Mode**: balanced (safe responses + refusals)
- **Training Type**: Supervised Fine-Tuning (SFT) for safety

## Safety Features
This model has been trained to:
- Provide helpful responses to safe prompts
- Refuse to engage with unsafe or harmful requests  
- Maintain safety boundaries while being helpful

## Usage
```python
from transformers import AutoTokenizer, AutoModelForCausalLM

model = AutoModelForCausalLM.from_pretrained("ybkim95/google_gemma-7b-it_safety_sft")
tokenizer = AutoTokenizer.from_pretrained("ybkim95/google_gemma-7b-it_safety_sft")

# Example usage
prompt = "User: [Your prompt here]\nAssistant:"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_length=200)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
```

## Model Files
This is a sharded model due to its size. All shards will be downloaded automatically when loading.