English prompt-continuation toxicity benchmark: models continue web-text prefixes; PerspectiveAPI scores how toxic the completions are.
unassessed
| Category | safety |
|---|---|
| Subcategory | continue English web-text prefixes and score PerspectiveAPI toxicity of the generations |
| Page status | active |
| Metric | expected_max_toxicity / max_toxicity_probability (paper); HELM also toxic_frac; OpenCompass toxic_frac_valid |
| Direction | lower_is_better |
| Dataset size | 99016 |
| Dataset licence | Apache-2.0 |
| Publisher | Allen Institute for AI / University of Washington |
RealToxicityPrompts (Gehman et al., Findings of EMNLP 2020) tests neural toxic degeneration: whether a language model continues an English web-text prefix into text that PerspectiveAPI marks toxic (TOXICITY ≥ 0.5). Prompts are the first half of OpenWebText sentences, stratified across four toxicity bins. Non-toxic prefixes still induce toxic continuations in the paper. HELM's real_toxicity_prompts scenario loads the full prompt file. OpenCompass's realtoxicprompts configs default to the challenging subset (prompts that elicited high toxicity from several 2020-era models). This page is that Gehman evaluation, not [decodingtrust_toxicity_prompts](decodingtrust_toxicity_prompts.md), which wraps a different DecodingTrust file set.
Open-ended English continuation. Paper: temperature 1, nucleus p=0.9, up to 20 tokens, k=25 generations. HELM completion adapter: temperature 0.9, max_tokens 20, num_outputs 5, no nucleus sampling (commented in classic_run_specs.py). OpenCompass: zero-shot GenInferencer on prompt_text, ToxicEvaluator via PerspectiveAPI (key ENV / PerspectiveAPIkey).
No model card in ModelSpec reports this benchmark yet.