Overview#

The public research leaderboard records refusal and compliance behavior for evaluated models. It reports uncensored rate and average compliance score from 0 to 1. These numbers describe benchmark behavior only: they are not a safety certification, a harm measurement, or proof of model identity.

Metrics#

uncensored_rate is the share of prompts scored as compliant rather than refusing.
avg_compliance_score is the mean evaluator score from 0 to 1.
Results are comparable only when corpus version, generation settings, and evaluator are identical.
Public submissions are unverified research claims unless the submitter also publishes reproducible artifacts.

Submit a result#

Install the package and run the benchmark with an explicitly selected evaluator before submitting.

pip install uncensorbench

python

from uncensorbench import UncensorBench, KeywordEvaluator
from uncensorbench.leaderboard import Leaderboard

bench = UncensorBench()
evaluator = KeywordEvaluator()
# Run the model over the chosen corpus and retain the result artifact.

Leaderboard(token="hf_write_token").submit({
    "model": "organization/model-revision",
    "uncensored_rate": 0.0,
    "avg_compliance_score": 0.0,
    "total_prompts": len(bench.prompts()),
})

Publication boundary#

A submission requires a write-scoped Hugging Face token. The client does not verify model identity, pin a benchmark version, redact responses, or provide an approval workflow. Review every field and linked response artifact before publication. Never upload generated harmful content, credentials, private prompts, or unredacted result files.