license: mit
task_categories:
- text-generation
tags: - benchmark
- llm
- abliteration
- censorship
- evaluation
- leaderboard
pretty_name: UncensorBench Leaderboard
UncensorBench Leaderboard
Public leaderboard for UncensorBench - a benchmark for measuring LLM censorship removal effectiveness.
About
This leaderboard tracks model performance on the UncensorBench benchmark, which evaluates how well models respond to prompts that typically trigger refusal behaviors.
Metrics
- Uncensored Rate: Percentage of prompts where the model provided a compliant response
- Average Compliance Score: Mean compliance score across all prompts (0-1 scale)
How to Submit
Install the package and run the benchmark:
pip install uncensorbench
from uncensorbench import UncensorBench, SemanticEvaluator, Leaderboard
# Run benchmark on your model
bench = UncensorBench()
evaluator = SemanticEvaluator()
# ... generate responses and evaluate ...
# Submit to leaderboard
leaderboard = Leaderboard(token="hf_your_token")
leaderboard.submit({
"model": "your-model-name",
"uncensored_rate": 0.15,
"avg_compliance_score": 0.23,
"total_prompts": 150,
...
})
Or use the provided notebook: establish_baseline.ipynb
Disclaimer
This benchmark is for research purposes only. Results should be interpreted in the context of AI safety research.