birgermoell/reasoning-traces-multilingual

OpenEuroLLM Multilingual Mathematical Reasoning Traces — TranslateGemma Pilot

Release status: private `v0.1-pilot` staging dataset. The rows passed the automated translation gates described below, but this pilot has not yet completed a systematic native-speaker audit or independent downstream-solver verification. It should not yet be treated as a final production training release.

This dataset contains 2,930 accepted translations of mathematical reasoning traces into 37 non-English languages used by OpenEuroLLM. Each row preserves an already-solved source problem, step-by-step <think> trace, and final answer. The translation model was instructed to translate faithfully rather than solve the problem again.

The approach follows the language-specific reasoning workflow described in AMD's RL training playbook: translate existing reasoning trajectories so that problems and solutions remain aligned across languages.

Release summary

Rejected rows and error logs are not included in this dataset. The 99.4% gate pass rate among rows that reached the accepted/rejected quality decision (2,930 of 2,947) should not be confused with the overall 79.2% accepted yield across all 3,700 attempted pairs.

Source data and sampling

The source is the SFT/math portion of `nvidia/Llama-Nemotron-Post-Training-Dataset`, version v1.1 in the retained row metadata. The pilot selected 100 deterministic reservoir samples from the first 20,000 streamed math rows. It uses 20 sources from each source-reasoning-length stratum:

• —fewer than 2,000 characters;
• —2,000–5,000 characters;
• —5,000–10,000 characters;
• —10,000–20,000 characters;
• —20,000 characters or more.

Every accepted row retains the upstream dataset, subset, split, source ID, source license, and a lossless serialized copy of the upstream metadata. The same 100 source IDs were attempted in each target language, which makes the pilot useful for aligned multilingual comparisons. English passthrough rows are not part of this release.

Translation procedure

• —Model: translategemma:27b (TranslateGemma 27B, Q4KM Ollama build)
• —Method: faithful_trace_translation
• —Hardware: two NVIDIA L4 GPUs, one Ollama worker per GPU
• —Decoding: temperature 0.0
• —Context / generation cap: 4,096 tokens / 4,096 tokens
• —Initial chunk limits: 4,500 source characters and 64 protected spans
• —Chunking: semantic splitting with adaptive subdivision on invariant failure
• —Protection: formulas, numbers, code, URLs, assignments, and structural markers were masked during translation and required to return exactly once

Long documents were translated in chunks and reassembled. The accepted release contains 24,889 translated chunks, averaging 8.446 chunks per document. Runtime fields are retained for reproducibility; duration values in translation_metrics are Ollama nanoseconds.

Automated quality gates

An output was accepted only when it had:

1. 1.non-empty translated prompt, reasoning trace, and final answer;
2. 2.plausible target-language confidence and expected writing-system evidence;
3. 3.no disallowed unexpected script relative to the source;
4. 4.exact protected-placeholder recovery during every translation chunk;
5. 5.preserved numbers, formulas, code, URLs, assignments, and section structure;
6. 6.plausible translated/source length ratio; and
7. 7.no high source-copy similarity indicating an untranslated output.

The 17 deterministic rejections were caused by 15 unexpected-script flags, one protected-token loss, and one target-language-confidence failure. No accepted row crossed the pilot's likely-untranslated threshold. These checks detect structural failures; they do not prove that every intermediate mathematical statement is semantically equivalent to the source.

Per-language outcomes

Each language had 100 attempted source traces. Error is the number of pairs that never produced an accepted or deterministically rejected row after retries.

Dataset structure

The dataset has one train split. It is conversational SFT data with a translated user problem and an assistant response containing the translated <think> trace followed by the final answer.

from datasets import load_dataset

dataset = load_dataset(
    "openeurollm/reasoning-traces-multilingual",
    split="train",
)

swedish = dataset.filter(lambda row: row["language"] == "sv")
print(swedish[0]["messages"])

Fields

Licensing and attribution

All 2,930 accepted pilot rows carry source_license: cc-by-4.0, so this pilot is marked CC BY 4.0. Users must preserve attribution and comply with the upstream Llama-Nemotron dataset card. The translation-pipeline software is Apache-2.0; that software license does not relicense dataset content. Future expansions must continue to expose licensing per row rather than assuming this pilot's uniform license distribution.

Intended uses

• —multilingual mathematical-reasoning SFT experiments;
• —cross-lingual reasoning and representation studies;
• —controlled language-mixture and ablation experiments;
• —evaluation of translation-based reasoning-data generation; and
• —pilot-scale downstream-solver comparisons before larger campaigns.

Maintain a strong high-quality English component in training mixtures and independently evaluate mathematical correctness, language quality, and reasoning-language consistency.

Limitations and risks

• —This is a small stratified pilot, not a representative sample of all mathematical reasoning.
• —The traces are synthetic source traces translated by a model, not human-authored explanations.
• —Accepted rows passed automated structural checks but not a systematic native-speaker audit.
• —Independent downstream-solver verification has not yet been completed.
• —A preserved final answer does not prove every translated intermediate claim is correct.
• —Low-resource and closely related languages remain difficult for automatic language identification.
• —Model translation can introduce unnatural phrasing, source-language leakage, or semantic drift.
• —Explicit reasoning traces can reproduce undesirable source reasoning patterns and should not be treated as factual references.
• —Yield varied by language; absent source/language pairs are not missing at random because failures were frequently caused by protected-placeholder handling on long traces.
• —Runtime host values describe the internal pilot workers and are not stable public endpoints.

Validation still required before public release

1. 1.A stratified native-speaker review across language families and reasoning-length strata.
2. 2.Independent answer and solver-accuracy verification on a fixed multilingual evaluation sample.
3. 3.Semantic equivalence checks focused on long, multi-chunk traces.
4. 4.Review of low-yield languages and systematic retry/fallback-model adjudication.
5. 5.PII and broader safety screening appropriate to the intended training use.

Reproducibility

The translation design, launch settings, and analysis scripts are maintained in the OpenEuroLLM reasoning-trace translation project. The pilot is append-only and resumable; accepted and rejected rows are stable, while failed source/language pairs can be retried with a fallback translator in a future version. A later release should use a new version tag and must not silently replace accepted text from this snapshot.

Citation

@misc{openeurollm_reasoning_traces_translategemma_pilot_2026,
  title        = {OpenEuroLLM Multilingual Mathematical Reasoning Traces: TranslateGemma Pilot},
  author       = {{OpenEuroLLM contributors}},
  year         = {2026},
  howpublished = {Hugging Face dataset},
  note         = {Private v0.1-pilot staging release}
}

Please also cite the upstream Llama-Nemotron dataset and any source papers requested by its dataset card.