k-chirkunov/jeeran_claims_comparison_generated

Jeeran Claims Comparison (Generated)

Synthetic dataset of Arabic claim pairs (short review "key points") labelled with the semantic relationship between claim_a and claim_b, across 7 categories: Equivalent, Contradictory, OverlapIncomparable, DifferentFacts, Uncertain, A-MoreSpecific, B-MoreSpecific.

Claims are written in Modern Standard Arabic or one of five dialects (Egyptian, Gulf, Maghrebi, Levantine, Saudi), span 17 business/service domains, and ~25% contain negation. Some claims carry realistic human typos (see error_a / error_b).

Generation used Gemini 2.5 Pro and GPT 5.4 (with thinking/reasoning). Each record was then independently re-classified by the other model; only records where the judge agreed with the intended label are kept, and each carries the judge's English reasoning trace in thinking.

Splits

train / test / eval (0.7 / 0.2 / 0.1), stratified by subcategory.

Columns

• id: incremental row id within the split (0-based counter)
• claim_a, claim_b: the two Arabic claims
• category: relationship label (one of the 7 categories)
• subcategory: finer relationship type, or category-only
• dialect: MSA / Egyptian / Gulf / Maghrebi / Levantine / Saudi
• has_negation: whether a claim contains negation
• error_a, error_b: injected typo class per claim (none if clean)
• flipped: whether the pair is a flipped (claim_a<->claim_b) mirror
• thinking: the judge model's English reasoning trace (non-empty)