README.md

metadata

license: apache-2.0
language:
  - en
  - zh
  - es
  - th
task_categories:
  - text-generation
tags:
  - math
  - mathematical-reasoning
  - multilingual
  - machine-translated
  - gsm8k
  - openr1
  - m-goal
size_categories:
  - 10K<n<100K
extra_gated_heading: Request access to Multilingual-Math-20K
extra_gated_prompt: >-
  Access to this dataset 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, (2) respect the upstream licenses of GSM8K (MIT)
  and OpenR1-Math-220k (Apache-2.0), and (3) not redistribute the raw files
  outside your research group without permission.
extra_gated_fields:
  Name: text
  Affiliation: text
  Intended use:
    type: select
    options:
      - Research on multilingual reasoning
      - Model training / post-training
      - Evaluation / benchmarking
      - label: Other
        value: other
configs:
  - config_name: en
    data_files: mixed_20k_en.jsonl
  - config_name: zh
    data_files: mixed_20k_zh.jsonl
  - config_name: es
    data_files: mixed_20k_es.jsonl
  - config_name: th
    data_files: mixed_20k_th.jsonl

Multilingual-Math-20K (EN / ZH / ES / TH)

The 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).

19,759 math reasoning problems, fully parallel across four languages (English original + Chinese / Spanish / Thai translations), each with a step-by-step solution ending in a unified \boxed{<answer>} and a gold answer field. The paper's main results use the English file only (English-only post-training); the zh/es/th files support the translated-mixture ablations (Tables 5, 6 and 9). Companion dataset: mgoal-knowledge-mcq (non-math domain).

• 79,036 problem–language pairs (19,759 × 4), aligned by qid
• Solutions included for every language (not just questions)
• Machine-translated with a multi-stage quality-escalation pipeline + automatic and human verification
• Per-record provenance: source dataset, subset, and which translation route produced it

Composition

Source	Records	Notes
GSM8K (train)	7,473	grade-school word problems; #### N answers converted to \boxed{N}
OpenR1-Math-220k (sampled)	12,286	olympiads 9,322 · cn_contest 1,115 · aops_forum 1,016 · amc_aime 715 · inequalities 80 · number_theory 38

Answer types: numeric 16,682 · letter 2,718 · LaTeX 236 · other 123.

Data fields

One JSON object per line; file mixed_20k_<lang>.jsonl carries question_<lang> / solution_<lang>:

Field	Description
qid	Stable problem id, identical across the four files (parallel alignment key)
source / subset	gsm8k or openr1 + fine-grained subset
problem_type, openr1_gid, openr1_cc	OpenR1 provenance metadata (null for GSM8K)
answer	Gold final answer; for OpenR1 extracted from the last \boxed{} of the English solution, for GSM8K the original label
question_<lang>	Problem statement in the file's language
solution_<lang>	Step-by-step solution in the file's language, ending in \boxed{<answer>}
translation_method	Which pipeline route produced this record's translation (see below)

Translation pipeline

Primary translation by DeepSeek (flash route), with an automatic quality-escalation ladder for records failing language-identification or answer-consistency checks: chunked re-translation → v4pro (stronger model) → best_of_10 sampling → smart_r1 → final_5x → nllb fallback → human fixes.

Route	ZH	ES	TH
flash (primary)	17,884	19,195	13,768
chunked	711	185	4,110
v4pro	901	208	10
best_of_10	218	139	1,348
smart_r1 / final_5x / ds_after_gpt / gpt5_mini	19	14	405
nllb	6	2	78
human / human_orphan	20	16	40
escalated beyond primary	9.5%	2.9%	30.3%

The Thai column illustrates why "just translate it" is not free: ~30% of Thai records required escalation beyond the cheapest translator to pass verification.

Quality control (v2 cleaning pass)

• Language identification (fastText): question+solution language match ≥ 99.0% per language (en 99.65 / zh 99.0 / es 99.6 / th 99.7)
• Answer consistency: final \boxed{} matches the answer field in 100.00% of records per language
• Removed 131 records from v1: translation pollution such as chat-template fragments or "I will translate…" artifacts (128), no \boxed{} in the English solution (2), EN/translation answer mismatch (1)
• Answer-leakage scrub: trailing "Answer: X" hints removed from ~110 questions per language
• Format unification: every solution in every language ends with \boxed{<answer>}

Full machine-readable details in stats.json and cleaning_report.json.

Intended uses

Research on multilingual mathematical reasoning: SFT / RL post-training, studying response-language behavior (e.g., whether models reason in the query's language), translated-data ablations, and cross-lingual evaluation. The four-way parallel structure (same qid across languages) supports controlled comparisons of the same problem across languages.

Limitations

• Translations are machine-generated (with the QC above); residual translation noise exists, most likely in Thai.
• The OpenR1 portion skews toward competition-level problems; the GSM8K portion is grade-school level. Filter by source/subset if you need a uniform difficulty band.
• ~15% of answers are non-numeric (letter/LaTeX); filter by answer type for exact-match numeric evaluation.
• Only questions and solutions were translated; no cultural localization was performed (names, currencies, and contexts remain as in the originals).

Licensing and attribution

Released under Apache-2.0, consistent with the upstream licenses: GSM8K (MIT, OpenAI) and OpenR1-Math-220k (Apache-2.0, Hugging Face). If you use this dataset, please also credit the 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}
}