SFT-Math-Multilingual-43k

Long chain-of-thought math solutions, each written end to end in one of six languages

_{Share of responses whose reasoning is in the requested language, Qwen3-4B before and after fine-tuning on this dataset}

Released reasoning models often think in English whatever language the user writes in. This dataset is multilingual supervision for the reasoning itself: in every example the problem, the full <think> trace and the final solution share one language.

• Teach a model to reason in the user's language. Fine-tuning Qwen3-4B on it moves target-language reasoning from 39.8% to 99.4% across Chinese, Spanish, French, Arabic and Russian.
• Warm-start multilingual RL. It is the SFT stage before RL on DAPO-Math-Multilingual-6Lang.
• Study multilingual reasoning. One source distribution, six languages, every problem in exactly one of them.

At a glance

Examples	43,218 train · 240 test (held-out loss, 40 per language)
Languages	English 7,460 · Chinese 7,214 · Spanish 7,212 · French 7,146 · Arabic 6,976 · Russian 7,210
Length	median 12,038 characters per solution, reasoning included
Source	OpenR1-Math-220k, translated by Qwen3-14B and filtered
Trained on it	Qwen3-4B-SFT-Math-Multilingual
Sibling	SFT-Math-Multilingual-56k, a separately built 56K set used at 8B

Quick start

from datasets import load_dataset

ds = load_dataset("96kevinli29/SFT-Math-Multilingual-43k", split="train")
ex = ds[0]
print(ex["language"])                 # one of en, zh, es, fr, ar, ru
print(ex["messages"][0]["content"])   # the problem
print(ex["messages"][1]["content"])   # <think> … </think> + solution ending in \boxed{…}

Fields: messages (user problem, assistant solution), language, data_source, category.

How it was built

1. Source. English long chain-of-thought traces from OpenR1-Math-220k.
2. Assign. Each trace gets one of the six languages in round-robin order, so every problem appears in exactly one language.
3. Translate. English traces stay as they are. The rest are translated chunk by chunk by a locally served Qwen3-14B; formulas and figures are detached first and reattached afterwards, and <think> tags are kept.
4. Filter. A trace is dropped if a chunk was truncated or the \boxed answer changed, then by seven gates: one <think> block, a recoverable boxed answer, balanced $ and braces, a bounded length ratio, no repetition, and target-language fidelity.

Quality audit

	Coherence (1–5)	Target-language fidelity (1–5)	Reaches the boxed answer	Math preserved	No major English residue
Chinese	4.90	4.62	99.6%	100.0%	96.0%
Spanish	4.78	4.40	99.6%	100.0%	94.0%
French	4.81	4.32	99.6%	100.0%	95.6%
Arabic	4.47	3.54	99.6%	99.6%	84.0%
Russian	4.76	4.40	100.0%	98.8%	94.4%
English, untranslated	4.88	n/a	100.0%	100.0%	n/a

250 sampled traces per language from the 43K corpus, scored by an independent judge.

Limitations

• Machine-translated. A trace can contain a stray word in English or another language; Arabic is the weakest language in the audit.
• Answers are not verified. Translation only checks that the boxed answer is unchanged from the source.
• Fine-tuning side effects. Models trained on it lose some accuracy and loop on about a fifth of non-English prompts until an RL stage removes the looping; see Qwen3-4B-SFT-Math-Multilingual.

Citation

@article{li2026posttraining,
  title   = {What Does Post-Training Change in Multilingual Reasoning?},
  author  = {Li, Hongyang and Li, Xiao and Wu, Caesar and Danoy, Gr{\'e}goire and Bouvry, Pascal},
  journal = {arXiv preprint arXiv:2609.37104},
  year    = {2026}
}

@misc{sft-math-multilingual-43k-2026,
  title        = {{SFT-Math-Multilingual-43k}: long chain-of-thought math solutions in six languages},
  author       = {Li, Hongyang and Li, Xiao and Wu, Caesar and Danoy, Gr{\'e}goire and Bouvry, Pascal and {Sea-Fill Community}},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/96kevinli29/SFT-Math-Multilingual-43k}},
  note         = {Maintained by Sea-Fill Community.}
}

Built and released by the authors together with the Sea-Fill Community, which maintains this release.

Please also cite OpenR1-Math-220k, the source of the English traces.