MultiverseComputingCAI/safety-for-whom

Safety for Whom? — Narrow-Boundary Safety Prompts (Politics & Religion)

This dataset accompanies the paper "Safety for Whom? Boundary-Aware Self-Distillation for Controlled LLM Safety Refusal" (arXiv:2609.04482, López-Ávila et al., 2026). See also the project write-up: "Safety for Whom? Refusing the Right Subset of a Topic, Not the Whole Topic".

The paper studies narrow-boundary safety: instead of asking "is this topic harmful?", it asks "which subset of a topic is incompatible with a given deployment policy?" A civics tutor and a public-sector assistant may share a base model but need different boundaries on the same topic of politics, refusing targeted manipulation while still answering factual questions. This release contains the prompts used to study that question, with no model-generated responses and nothing tied to a particular target model, so that anyone can build their own training or evaluation pipeline on top of the same prompt pool.

Considerate use

Many of these prompts describe or request harmful political or religious manipulation, propaganda, or persuasion by design, since that is what a boundary-safety benchmark needs to measure. Please use this data for safety alignment, refusal-boundary evaluation, and related research, not to build persuasion or manipulation capabilities. The political and religious framings named in the prompts are there for construction purposes and do not represent any endorsement or judgement by the authors.

What this is (and isn't)

This is a prompts-only release. In the paper, these prompts are paired with model-generated, WildGuard-filtered responses (refusal traces for harmful prompts, compliance traces for benign ones); that response layer is specific to the target model and training approach (the paper's main results use Qwen3-8B) and is not included here. What's included is the model-independent layer: the prompts themselves, plus the metadata used for the paper's stratified evaluation (persona, topic, and length-bucket labels).

Two topical domains are covered:

• Politics — the paper's main experimental domain. Political persuasion is a useful testbed because manipulative persuasion is harmful while factual political information stays legitimate, which is exactly the case where topic-level refusal is too blunt.
• Religion — built with the same generation pipeline (same four-stage hierarchical prompting, same pairwise-boundary construction). Religion has its own results in the paper's appendix, showing the same boundary-safety pattern holds on a second topic.

Files

Each domain has three parts: main (single harmful/target prompts), pairwise (boundary pairs), and test (held-out evaluation sets).

Politics

Role	File	Rows	Notes
Main	Politics_new_train.json	40,480	Harmful political prompts from the topic→subtopic→point→persona pipeline.
Pairwise	Politics_new_train_pairwise.json	40,453	Boundary pairs: prompt_a (comply-expected) / prompt_b (refuse-expected) sharing a topic anchor.
Test — in-distribution	politics_test_in.json	6,000	Held-out prompts from the same generation process as Main.
Test — out-of-distribution	politics_test_out.json	1,540	A separate construction run from an earlier version of the pipeline, before prompt-length control was added. Not a paraphrase of the in-distribution set; its subtopics differ.
Test — boundary, refuse-side	pol_pr_ood_test.json	1,539	Held-out prompt_b boundary prompts, sourced from politics_test_out.
Test — boundary, comply-side	pol_pb_ood_test.json	1,539	Held-out prompt_a boundary prompts, from the same pairs.

The 1,539-per-side boundary count matches the paper's reported figure for its held-out harmful-benign pairs, so these two files are exactly the paper's headline boundary-evaluation set. The paper's training data also includes a labeled, coverage-repaired version of the harmful prompts (40,293 rows after retry-based repair); that labeled training file is a downstream, approach-specific artifact and isn't part of this prompts-only release.

Religion

Role	File	Rows	Notes
Main	Religion_new_train.json	40,675	Same pipeline, religion-specific subtopic and persona guidance.
Pairwise	Religion_new_train_pairwise.json	40,643	Same boundary-pair construction as politics.
Test — in-distribution	religion_test_in.json	195
Test — out-of-distribution	religion_test_out.json	152
Test — boundary, refuse-side	rel_pr_ood_test.json	152	Sourced from religion_test_out.
Test — boundary, comply-side	rel_pb_ood_test.json	152	From the same pairs.

Religion's held-out evaluation sets are smaller than politics' (152–195 rows versus 1,539–6,000). This reflects the smaller scope religion was originally built at; the paper's religion results are computed on these same held-out sets.

Schema

The main-pool files differ slightly between domains:

• Politics_new_train.json: _id, length_range, persona, point, prompt, subject, subtopic
• Religion_new_train.json: _id, instruction, length_range, persona, persona_bucket, point, prompt, prompt_label, subject, subtopic — religion's pipeline additionally pre-computes persona_bucket and prompt_label, and carries the prompt text under both instruction and prompt.

Pairwise files (*_pairwise.json), both domains: dataset, generation_idx, metadata, pair_id, prompt_a, prompt_b, task, with metadata a nested object carrying subject, subtopic, and point.

Test files:

• politics_test_in.json, politics_test_out.json: _id, instruction, length_range, persona, persona_bucket, point, prompt_label, subject, subtopic, prompt (prompt duplicates instruction).
• religion_test_in.json, religion_test_out.json: same fields as above.
• The four *_ood_test.json files carry the same fields as the in/out test files.

length_range takes one of 2-4 words, 4-8 words, 8-15 words, or more than 15 words, sampled per prompt at generation time. The two longer buckets are intentionally over-weighted to counteract the short-prompt bias of unconstrained LLM generation (see the paper for the rationale).

prompt_label takes one of vanilla_harmful, adversarial_harmful, vanilla_benign, or adversarial_benign, derived from the persona bucket that generated the prompt (moderate personas produce vanilla_*, radical personas produce adversarial_*). Every religion prompt_label is adversarial_harmful, because religion's persona pool has no moderate cluster; all five of its persona clusters are adversarial by construction. Politics mixes moderate and radical personas.

Politics train/test overlap audit

We audited the politics splits using the same normalised SequenceMatcher criterion and 0.86 threshold used by the generation pipeline. The audit compares unique subtopics exhaustively and then inspects prompts under every matched subtopic pair. We report both matched-pair density and test-side coverage: pair density alone can look artificially small because its denominator is the full Cartesian product.

Comparison	Matched subtopic pairs	Pair density	Test-side subtopic coverage
politics_test_in vs. politics_test_out	1 fuzzy; 0 exact	1/9,800 (0.0102%)	1/100 (1.00%) and 1/98 (1.02%)
each politics training artifact vs. pooled tests	8, including 3 exact	8/86,922 (0.0092%)	8/198 (4.04%)

The eight train/test candidates involve six politics_test_in subtopics and two politics_test_out subtopics. The response-filtered training artifact and the prompts-only source artifact produce the same eight candidate subtopic identities, so these are not counted as 16 independent overlaps. We then exhaustively compared every prompt pair inside all flagged subtopic pairs: 157,803 prompt pairs in total, of which zero reached the 0.86 threshold. The highest score was 0.800, and zero of the 7,540 held-out prompts had a confirmed lexical training match under this criterion. A separate three-pass semantic review conservatively identified three candidate subtopic groups with related requested information. Thus, the audit finds no lexical prompt near-duplicates while acknowledging limited reuse of political events, policy themes, and intent.

The pure prompt outputs from this audit can be found in the subfolder politics_audit.

Known limitations

• Longer prompts near the boundary. The boundary test sets skew heavily toward the more than 15 words bucket (politics: 95.3% on the refuse-side, 71.0% on the comply-side; religion: 91.4% and 49.3%), and politics_test_out is 98.6% in that bucket. This follows from how boundary pairs are built: they need enough shared context to encode a comply-versus-refuse distinction, and the out-of-distribution sets come from an earlier, pre-length-control run of the pipeline that tends to produce longer prompts on its own. This is expected behaviour, not a construction error.
• Persona labels on the four *_ood_test.json files mix personas from both domains. The persona and persona_bucket columns on pol_pr_ood_test, pol_pb_ood_test, rel_pr_ood_test, and rel_pb_ood_test draw from the full cross-domain persona pool rather than the correct per-domain pool. These four files are released exactly as used to compute the paper's results, so the dataset matches the paper's numbers precisely. This doesn't affect those numbers: the paper's reported PR-OOD/PB-OOD metrics are aggregate refusal rates over all rows in each file, and don't use persona or persona_bucket. If you run a per-persona-bucket breakdown on these four files specifically, treat that column as unreliable.
• politics_test_out.json only has two of politics' four persona buckets (RADICAL_RIGHT and RADICAL_LEFT, no MODERATE, no RADICAL_OTHERS) and no vanilla_harmful rows. This is an asymmetry worth knowing about if you compare it directly against politics_test_in.
• The overlap audit is hierarchical and lexical. All subtopic combinations are screened first; the full prompt cross-product is then evaluated only inside subtopic pairs meeting the 0.86 threshold. This checks 157,803 of 619,082,740 theoretical prompt pairs (0.025490%). The audit establishes zero lexical near-duplicates under the pipeline criterion, not complete semantic independence. Shared subtopics should still be treated as potential semantic overlap when their prompt wording differs.

Citation

If you use this dataset, please cite the paper:

@article{lopezavila2026safety,
  title   = {Safety for Whom? Boundary-Aware Self-Distillation for Controlled LLM Safety Refusal},
  author  = {L{\'o}pez-{\'A}vila, Alejo and Garc{\'i}a-Ferrero, Iker and Garcia, Jezabel and Tiene, Antonio and Or{\'u}s, Rom{\'a}n},
  journal = {arXiv preprint arXiv:2609.04482},
  year    = {2026},
  url     = {https://arxiv.org/abs/2609.04482}
}

Contact

Multiverse Computing — multiversecomputing.com