Chia-Mu-Lab/reasoning-seed-6k-sqa-pqa-hqa

reasoning-seed-6k (StrategyQA + ProntoQA + HotpotQA)

Seed question+answer pool for reasoning-trace distillation. CoT is meant to be generated later by a strong teacher model from each question; no gold CoT is included in question.

6,000 rows = 2,000 each, sampled with random.Random(7).

source	reasoning type	n	answer_type	answer example
strategyqa	commonsense (implicit multi-step)	2000	yes_no	Yes / No
prontoqa	symbolic (first-order-logic, fictional ontology)	2000	true_false	True / False
hotpotqa	multi-hop QA (distractor, all hard)	2000	span_or_yesno	span or yes/no

Fields

• id, source
• question — self-contained prompt to feed the teacher (HotpotQA folds in its 10-paragraph context; ProntoQA folds in premises)
• answer — gold answer string
• answer_type — yes_no | true_false | span_or_yesno
• meta_json — per-source extras (StrategyQA qid/term; HotpotQA type/level/supporting_titles)

Sources

• StrategyQA — ChilleD/StrategyQA (official 2,290 answered train pool, re-split; answers are yes/no)
• ProntoQA — longface/prontoqa-train (parsed to premises+query / True-False; fictional ontology)
• HotpotQA — hotpotqa/hotpot_qa config distractor, validation split (all difficulty=hard)

Scoring guidance

• strategyqa / prontoqa → binary exact-match
• hotpotqa → official EM + F1 with normalization (must handle yes/no as well as spans)