Measure over-refusal in LLMs using OR-Bench
Run benchmarks on prediction models
Explore and visualize diverse models
Browse and submit LLM evaluations
Evaluate and submit AI model results for Frugal AI Challenge
Browse and submit model evaluations in LLM benchmarks
Analyze model errors with interactive pages
Track, rank and evaluate open LLMs and chatbots
Convert Stable Diffusion checkpoint to Diffusers and open a PR
Calculate survival probability based on passenger details
Evaluate adversarial robustness using generative models
Convert and upload model files for Stable Diffusion
Launch web-based model application
OR-Bench Leaderboard is a tool designed to measure and benchmark over-refusal in Large Language Models (LLMs). It provides a standardized framework to evaluate when and how models refuse to answer questions or generate text. This leaderboard helps researchers and developers understand the limitations and safety mechanisms of LLMs by comparing their performance across different scenarios.
• Model Comparison: Allows users to compare multiple models based on their refusal patterns.
• Refusal Trigger Evaluation: Tests models against a curated set of triggers to assess their refusal thresholds.
• Metric Aggregation: Provides aggregated metrics such as refusal rates and response patterns.
• Result Sharing: Enables sharing of benchmark results for community collaboration.
• Documented Methodology: Offers transparent documentation of evaluation methods and datasets.
1. What is the purpose of OR-Bench Leaderboard?
The purpose of OR-Bench Leaderboard is to provide a standardized way to measure and compare over-refusal behaviors in LLMs, helping to identify models with balanced safety and utility.
2. Why is measuring over-refusal important?
Measuring over-refusal is important because it helps assess a model's ability to avoid harmful or inappropriate responses while still providing useful answers.
3. How can I interpret the results from OR-Bench Leaderboard?
Results show how often a model refuses to answer and under what conditions. Lower refusal rates may indicate a model that is more permissive, while higher rates suggest stricter safety mechanisms.