Initial release

README.md

@@ -0,0 +1,106 @@
+---
+license: apache-2.0
+language:
+- en
+- zh
+- es
+- fr
+- ar
+- ru
+task_categories:
+- text-generation
+tags:
+- math
+- reasoning
+- multilingual
+- multilingual-reasoning
+- chain-of-thought
+- supervised-fine-tuning
+- competition-math
+size_categories:
+- 10K<n<100K
+source_datasets:
+- open-r1/OpenR1-Math-220k
+configs:
+- config_name: default
+  data_files:
+  - split: train
+    path: train.parquet
+  - split: test
+    path: test.parquet
+---
+
+# SFT-Math-Multilingual-56k: long chain-of-thought math traces in six languages
+
+**56,307 mathematics problems, each with a full chain-of-thought solution written in one of six languages** (English, Chinese, Spanish, French, Arabic, Russian). The problem and the whole reasoning trace share one language, so a model fine-tuned on this corpus learns to reason in the language of the problem instead of switching to English.
+
+This is the supervised fine-tuning corpus behind [`96kevinli29/Qwen3-8B-SFT-Math-Multilingual`](https://huggingface.co/96kevinli29/Qwen3-8B-SFT-Math-Multilingual) (Qwen3-8B, epoch 1) in the paper *What Does Post-Training Change in Multilingual Reasoning?* ([arXiv:2609.37104](https://arxiv.org/abs/2609.37104)). It was constructed separately from the 43,218-trace corpus used at 4B ([`96kevinli29/SFT-Math-Multilingual-43k`](https://huggingface.co/datasets/96kevinli29/SFT-Math-Multilingual-43k)), with the same pipeline.
+
+## Contents
+
+| Language | train | test (held-out loss) |
+| :--- | ---: | ---: |
+| English (`en`) | 9,760 | 40 |
+| Chinese (`zh`) | 9,294 | 40 |
+| Spanish (`es`) | 9,451 | 40 |
+| French (`fr`) | 9,357 | 40 |
+| Arabic (`ar`) | 9,306 | 40 |
+| Russian (`ru`) | 9,139 | 40 |
+| **Total** | **56,307** | **240** |
+
+Every source problem appears in exactly one language.
+
+## Schema
+
+| Field | Description |
+| :--- | :--- |
+| `messages` | chat list: `[{"role": "user", "content": <problem>}, {"role": "assistant", "content": "<think> … </think> <solution with \\boxed{answer}>"}]` |
+| `data_source` | `"openr1_math"` |
+| `category` | `"math_competition"` |
+| `language` | one of `en, zh, es, fr, ar, ru` |
+
+```python
+from datasets import load_dataset
+ds = load_dataset("96kevinli29/SFT-Math-Multilingual-56k")
+print(ds["train"][0]["language"], ds["train"][0]["messages"][1]["content"][:300])
+```
+
+## How it was built
+
+1. **Source.** English long chain-of-thought traces from [`open-r1/OpenR1-Math-220k`](https://huggingface.co/datasets/open-r1/OpenR1-Math-220k).
+2. **Assign.** Each trace is assigned one of the six languages in round-robin order by its index.
+3. **Translate.** English-assigned traces are kept unchanged. The others are translated by a locally served Qwen3-14B, chunk by chunk: formulas, figures and display environments are detached before translation and reattached afterwards, and the `<think>` tags are preserved. (Translating a long trace in one pass makes the translator re-solve the problem instead of translating it.)
+4. **Filter.** A trace is dropped if any chunk was truncated or the `\boxed` answer changed, and then by seven further gates: exactly one `<think>` and one `</think>`; a recoverable boxed answer; an even number of dollar signs; matched braces; a bounded translated-to-source length ratio; no repetition (zlib compresses the trace by more than 70%); and target-language fidelity (share of prose characters in the target script above 0.5 for Chinese, Arabic and Russian; share of English function words below 0.15 for Spanish and French).
+
+## Quality audit
+
+An audit of 250 sampled traces per language from the 43K corpus. Coherence and target-language fidelity are independent judge scores on a five-point scale; the last column is the share of traces free of moderate or severe English residue. The English row is the untranslated source, the reference these scores are read against.
+
+| Language | Coherence | Target-language fidelity | 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 |
+
+## Limitations
+
+- **Machine translation.** Translated traces can contain a stray word in English or in another language (for example a discourse marker such as "Okay," at the start of the reasoning). Arabic is the weakest language in the audit.
+- **Answers are not verified.** The source traces were filtered structurally, not for answer correctness; the translation step only checks that the boxed answer is unchanged.
+- **What fine-tuning on it does.** In the paper, models fine-tuned on this corpus reason in the requested language almost always, but lose some accuracy and loop on about a fifth of non-English prompts until an RL stage removes the looping. See the model cards for numbers.
+- Mathematics only.
+
+## Citation
+
+```bibtex
+@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}
+}
+```
+
+Please also cite [OpenR1-Math-220k](https://huggingface.co/datasets/open-r1/OpenR1-Math-220k), the source of the English traces.

test.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cd4700538d6545352702ceff6c4d7c38e298ad06b8b8b5a2564582ed2fe121f7
+size 1933095

train.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:90aec9221220d24188661ed022c179b4f4a90b0468641f090c261f4e1596c9ac
+size 478058444