LENS-WarBias

Version 1.3 — research draft; independent human validation pending.

LENS-WarBias is a Ukrainian–English prompt dataset for studying war-related stereotype elicitation and transfer after model unlearning. It covers 981 WarBias matrix entries, 129 case families, 15 actor profiles and 56 actor/gender/age variants. It contains prompts and provenance metadata, not target-model responses, a validated forget set, or measured model scores.

The dataset contains deliberately leading stereotype probes for research. Their premises are claims under test, not factual assertions endorsed by the authors. English and Ukrainian versions are intended to be semantically paired; independent bilingual validation remains pending.

Languages and dataset layout

Choose the uk or en subset in the dataset viewer. Each uses the same schema with a single prompt column.

Subset	Language	Rows
uk (default)	Ukrainian	3,201
en	English	3,201

Each row includes prompt_id, unique_pair_id, language, level, probe_type, prompt, case_partition, usage_role and split_group_id. Match translations across subsets using unique_pair_id; prompt_id identifies a specific language version. The subsets contain exactly the same paired probe families.

There are 6,402 unique language-specific prompts. The downloadable matrix-linked exports cover all 981 source entries with five probe pairs each: 4,905 bilingual pairs, equivalent to 9,810 language slots before deduplication. Abstract prompts are shared across actor variants; matrix_prompt_mapping.csv preserves all links. Paired-column exports remain available for provenance and alignment, but are not the viewer's language subsets. Do not concatenate these alternative exports as independent examples.

The Hub split name unassigned means that train/development/test membership has not been determined. It is not a training split; language is selected using the subset name.

Prompt levels

Level / probe	Purpose	Matrix-linked pairs	Unique pairs
L0 / direct	Direct, case-specific question about the target association	981	981
L1 / attributed	Third-party claim and a question about its reasoning	981	981
L2 / contrastive	Comparison with a short, relevant counterpoint	981	981
L3 / abstract_reasoning	Explanation of an abstract situation with actors masked	981	129
L3 / group_identification	Explicit request to identify group A in the same situation	981	129

L0 is deliberately leading and does not measure spontaneous bias under neutral prompting. L1 and L2 probe transfer under attribution and comparison. L3 reconstructs an abstract situation rather than merely quoting a claim with a masked name. Ask the two L3 forms in separate conversations. Explicit identification must be analysed separately from spontaneous identity recovery.

The source actor behind group A is provenance, not a uniquely correct identification answer. Insufficient information is a valid response. Identification, endorsement, attributed description, rejection, uncertainty, refusal and degraded generation require separate labels.

Read bilingual examples.

Loading

from datasets import load_dataset

uk = load_dataset("FairForget/LENS-WarBias", "uk", split="unassigned")
en = load_dataset("FairForget/LENS-WarBias", "en", split="unassigned")

uk_prompt = uk[0]["prompt"]
en_prompt = en[0]["prompt"]
assert uk[0]["unique_pair_id"] == en[0]["unique_pair_id"]

For reproducibility, pin the Hub commit using revision=. Match translations by ID instead of relying on row order.

Send only the prompt text to the respondent model. Do not include hidden actor metadata, association labels, source evidence IDs or masking notes in model input. Do not use the dataset-authoring instructions as a respondent system message. This release contains no model responses.

Construction and language policy

The assistant authored case-level English claim translations, bilingual alternatives, direct questions, abstract situations and actor terminology. A deterministic renderer applied these resources across the frozen matrix. These are assistant-authored, template-rendered prompts, not 9,810 independent sampled generations. No external generation API was used to create this LENS release. The separate Lapa triplet run is not included.

Use familiar, context-appropriate terminology in each language while matching the referent. For example, Ukrainian «ВПО» pairs with English “internally displaced people”; «ПТСР» pairs with “PTSD”. Abbreviation symmetry is not required. These are editorial choices, not corpus-validated frequency findings. See terminology choices and policy. Differences in wording remain a possible influence on cross-language results.

Prompts have no embedded answer-length instruction. Wartime openings do not specify a country by name; country names remain where integral to the source actor or association. Host-country settings distinguish the country of origin from the place where the interaction occurs.

Validation and limitations

Structural checks cover source hashes, IDs, five-probe coverage, paired languages, placeholder replacement, English script, matching L3 premises, explicit actor masking, direct-question structure, selected quantifier markers and export consistency. A portable rebuild reproduced six core artifacts byte for byte. See quality report and validation results.

These checks do not establish semantic validity, bilingual equivalence, equal difficulty across levels or elicitation effectiveness. Author review is not independent human validation. All revised families require review; the priority queue contains 44 items. Abstraction can reduce specificity, including military-to-civilian skills becoming transfer between fields B and C, and motherhood becoming parenthood. One direct question interprets the source occupation/staying relation as an inference to political support; this is explicitly flagged.

Actor profiles and case families can overlap. The 127 proposed split_group_id groups reflect case identity, exact claim overlap and duplicate prompts; semantic-overlap review must precede split allocation. Keep related actor variants and both languages together. Distinguish transfer on trained associations from generalisation to held-out cases.

Only reviewed Ukrainian L0 responses from training cases are candidates for the proposed own-response unlearning condition. A generated answer is not automatically an endorsed stereotype or an eligible forget example. The separate triplet benchmark and shared forget dataset remain Ukrainian and are not part of this upload.

Files and provenance

• data/uk.jsonl and data/en.jsonl: viewer-ready language subsets, each with one prompt field.
• Original CSV and JSONL exports for matrix linkage, paired alignment and combined-language analysis.
• matrix_prompt_mapping.csv, actor_terminology.csv and case_families.json for linkage and terminology.
• source_grammar_adjustments.csv for 18 source-derived grammar/punctuation adjustments.
• provenance/source/ for the frozen source matrix and references.
• provenance/authoring/ for the authored case resources.
• provenance/build_lens_warbias_bilingual.py for a standard-library-only rebuild. Run it with Python 3.10 or later; it writes a separate rebuild/ folder. provenance/build_language_subsets.py then derives the language subsets from its prompts_long.jsonl export.
• manifest.json and checksums.sha256 for versions and file hashes.

The repository includes the frozen source resources used for generation. Prior human-evaluation ratings, private annotation feedback, credentials and target-model responses are not included.

Attribution

The level-based design adapts Makovska and Fletcher, Between Suppression and Collapse: Evaluating Narrative Unlearning with LENS. WarBias case content, the explicit group-identification question, bilingual terminology policy and the shared-versus-own-response experiment are project-specific adaptations. This dataset has not yet been independently validated as a replication of that paper's protocol.

Distribution status

Public research draft distributed under Creative Commons Attribution 4.0 International. Attribute the LENS-WarBias dataset and its creators, link the licence, and indicate modifications. Referenced third-party materials retain their own terms; source references and attribution are retained in the provenance files.