jnm38/llm-safety-assessment

LLM Safety Evaluation Space

This Hugging Face Space allows you to evaluate the safety of Large Language Models (LLMs) by generating responses to prompts from safety datasets.

Features

• —Load and evaluate various LLMs (e.g., SmolLM, Llama)
• —Use safety datasets like AgentHarm, HH-RLHF, etc.
• —Customize sampling parameters (temperature, max tokens, etc.)
• —View results in JSON format

Usage

1. 1.Enter the model name (Hugging Face model ID).
2. 2.Specify the dataset name and optional config/split.
3. 3.Set the number of samples and generation parameters.
4. 4.Click "Run Evaluation" to generate responses.
5. 5.View the results showing prompts and model responses.

Requirements

• —GPU-enabled Space for vLLM compatibility.
• —HF_TOKEN secret for accessing private models/datasets (optional).

Note

This is a demo for evaluating LLM safety. Ensure compliance with dataset licenses and ethical guidelines.