ybkim95/gemma-3-4b-it_invthink

google/gemma-3-4b-it Safety SFT

This model is fine-tuned from google/gemma-3-4b-it using the Nvidia Aegis AI Content Safety Dataset 2.0.

Training Details

• Base Model: google/gemma-3-4b-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

from transformers import AutoTokenizer, AutoModelForCausalLM

model = AutoModelForCausalLM.from_pretrained("ybkim95/google_gemma-3-4b-it_safety_sft")
tokenizer = AutoTokenizer.from_pretrained("ybkim95/google_gemma-3-4b-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.