emolero/refusal-frontier

Observed Refusal Responses — Five Providers

1,000 short synthetic user prompts, submitted once to each of five model providers through OpenRouter. Assistant responses are observed API outputs preserved verbatim. The selected view contains one refusal per prompt, with 1,000 distinct normalized response texts.

Browse the data

Viewer configuration	Rows	Contents
selected (default)	1,000	One distinct refusal response per prompt
all_refusals	4,558	All complete responses classified as strict refusals
all_attempts	5,000	Every prompt/model cell, including safe redirections, mixed outputs, provider interventions, and errors
prompts	1,000	Synthetic user prompts and construction-review metadata

The split name data means the complete collection; no train/test split is supplied.

See 50 examples · Browse refusal categories · Download the complete release ZIP

The CSV files support the dataset viewer. Companion JSONL files preserve richer model provenance, annotation evidence, and review metadata. Category columns in CSV contain semicolon-separated labels.

Collection and models

Model requested	Hosting provider	Complete responses	Strict refusals	Selected
openai/gpt-6-sol	OpenAI	740	737	150
anthropic/claude-sonnet-5	Anthropic	961	955	238
google/gemini-3.8-flash	Google AI Studio	870	868	176
qwen/qwen3.8-max-0902	Alibaba	999	998	235
mistralai/mistral-medium-3-5	Mistral	1,000	1,000	201

Every response model received the same safety-focused system instruction explicitly asking it to decline harmful assistance and respond in one to four sentences. Routing was pinned with provider fallback disabled, and generation attempts were not automatically retried. The full instruction and model settings are in plan.json.

Of 5,000 attempted cells, 4,570 produced complete assistant responses: 4,558 classified strict refusals, five safe non-refusals, and seven mixed responses. Another 294 cells were provider safety interventions and 136 were technical errors. Text from mixed outcomes and provider interventions is withheld in this release. Such cells remain represented in the outcome metadata.

Refusal categories

Categories use three separate axes. Outcomes partition all 5,000 attempted cells. Stated-reason and response-feature categories apply to the 4,558 strict-refusal cells and can overlap.

• Stated reasons: harm/safety, rights/consent, deception/fraud, law, ethics/fairness, policy/rules, capability/access, other stated reason, or no stated reason.
• Response features: safe alternative, concrete safe example, clarification, empathy, consequence warning, conditional help, or support referral.
• Outcomes: strict refusal, safe non-refusal, mixed, harmful compliance, capability-only, uncertain, provider safety intervention, or technical error. Some outcome groups contain zero cases.

Each substantive reason/feature label in JSONL includes an exact evidence excerpt. Labels describe what a response states, not verified legal or medical facts or the model's internal motives. Merely naming a rejected action is not automatically a stated explanation. See taxonomy.json.

Construction, review, and limitations

The prompts comprise 25 categories × 10 authored intent families × four phrasings. Three assistant agents authored and cross-reviewed them. They are synthetic inputs, not real user logs. The prompt review board preserves proposed revisions and decisions.

Outcome labels and detailed refusal types were assigned by models, with partial assistant-agent review and recorded coordinator corrections. No human validation is claimed. There were 126 unique response-level peer reviews and 80 unique type-annotation peer reviews; the two review scopes should not be added as disjoint cohorts. Automatic labels remain fallible, including possible missing referrals and conditions or inferred reasons. All final type annotations pass format and verbatim-evidence checks, which do not establish semantic correctness.

This collection measures observed responses under a shared refusal-focused instruction. It does not estimate spontaneous refusal rates, internal honesty, general alignment, or real-world request prevalence. The 1,000 selected texts are distinct after normalization; that is not proof of semantic or stylistic independence. Some prompts are ambiguous or policy-dependent, and there is no benign control set.

Keep all four variants and every model response sharing a family_id together when creating data splits or grouped analyses. Cross-family semantic overlap can remain. Do not sum multi-label category counts as independent examples.

Full methodology and annotation limitations · Counts · Verification

Files and provenance

The response JSONL records contain prompt and family identifiers, requested and returned model/provider, generation IDs, timestamps, exact response hashes, and annotation evidence. Selected responses were checked against the saved API request and response records. Export reproducibility and release checks are documented in the verification file.

The raw API ledger, raw outcome dumps, standalone annotations spanning mixed outcomes, credentials, and console logs are excluded. Hashes of omitted private source artifacts appear in the release manifest. The included ZIP is the previously verified release; this repository adds the dataset card and upload manifest for Hugging Face.

The prompt files' model_tested: false fields refer to the earlier prompt-construction snapshot. This later response collection is linked by prompt ID and the frozen source-file hash.

No dataset license has been assigned in this upload. This dataset card does not grant additional rights in provider-generated content.

Viewer configuration follows Hugging Face's documented dataset configuration format.