Access is granted manually for research purposes. Please briefly describe your intended use. By requesting access you agree to (1) use the data for research only, and (2) respect the upstream licenses of ECQA (CDLA-Sharing-1.0) and CommonsenseQA (MIT).

M-GOAL Knowledge-MCQ (ECQA-derived, EN / ZH / ES / TH)

The knowledge-domain (non-math) post-training corpus from the paper "M-GOAL: Toward Accurate and Language-Consistent Reasoning in LLMs via English-Only Post-Training" (EMNLP 2026, to appear). 6,155 commonsense multiple-choice questions derived from ECQA, with chain-of-thought SFT targets in English and GRPO prompts in four languages, aligned by qid.

This corpus was used for the paper's cross-domain (ECQA → MMMLU) experiments, showing that the accuracy–language-consistency trade-off of English-only post-training — and its repair by adapter merging — is not specific to mathematics.

Files

File	Records	Content
`sft_en.jsonl`	6,155	English SFT text: MCQ question + options + instruction + chain-of-thought explanation ending in `\boxed{<letter>}`
`grpo_en.jsonl`	6,155	English GRPO prompts (question + options + instruction) with `answer_label`
`grpo_zh.jsonl` / `grpo_es.jsonl` / `grpo_th.jsonl`	6,155 each	The same items with question and options machine-translated into Chinese / Spanish / Thai (translated-mixture ablations)

Data fields

qid — stable id, identical across all five files (parallel alignment key)
lang — language of this record
text (SFT file) — full training text: question, options A–D(E), answer instruction (\boxed{<letter choice>} format), and the explanation
prompt (GRPO files) — question + options + answer instruction
answer_label — gold option letter

Construction

Questions and options come from ECQA (CommonsenseQA questions with human-written explanations). SFT targets convert ECQA explanations into a step-by-step rationale ending in a unified \boxed{<letter>}. Non-English prompt versions were produced with the same DeepSeek translation + fastText verification + escalation pipeline used for our math corpus (see companion dataset multilingual-math-20k).

Note: the paper's main M-GOAL knowledge results use English-only SFT/GRPO from this corpus; the zh/es/th prompt files exist for the translated-mixture dose-response ablations.

Licensing and attribution

Released under CDLA-Sharing-1.0, matching upstream ECQA (CDLA-Sharing-1.0; ECQA itself builds on CommonsenseQA, MIT). Please credit both upstream datasets.

Citation

@inproceedings{mgoal2026,
  title     = {M-GOAL: Toward Accurate and Language-Consistent Reasoning in LLMs via English-Only Post-Training},
  booktitle = {Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing (EMNLP)},
  year      = {2026},
  note      = {To appear}
}

Downloads last month: 6

Total file size:

21.7 MB