openeurollm/reasoning-traces-multilingual

OpenEuroLLM Multilingual Mathematical Reasoning Traces — Two-Stage Pilot

Release status: private `v0.2-pilot` staging dataset. All published rows passed the deterministic translation gates described below. This pilot has not yet completed a systematic native-speaker audit or independent downstream-solver verification and is not a final production training release.

This dataset contains 3,425 accepted translations sampled from 100 mathematical reasoning traces into 37 non-English OpenEuroLLM target languages. Accepted rows represent 99 source problems: the remaining source, nemotron-math-13867, had malformed source <think> tags and therefore produced no accepted translation in any target language. Each published row preserves an already-solved source problem, its step-by-step <think> trace, and final answer. Translation was performed faithfully; the translation models were instructed not to solve the problem again.

The release combines a primary TranslateGemma campaign with a Gemma 4 fallback pass over every source/language pair that did not pass the primary campaign. English passthrough rows are not included.

Release summary

Rejected rows and error logs are not included in the published train split. The 3,700 attempted pairs all have one final accepted, rejected, or error outcome. The final accepted yield is 92.6%, up from 79.2% in v0.1-pilot.

What changed in v0.2

• —Added 495 accepted fallback translations produced by gemma4:31b.
• —Increased accepted coverage from 2,930/3,700 to 3,425/3,700.
• —Preserved every row from v0.1-pilot; fallback output was appended only for previously absent source/language pairs.
• —Recomputed per-language outcomes and dataset metadata.
• —Added release/v0.2-pilot.json with machine-readable counts, provenance, and checksum.

Source data and sampling

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

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

Every published 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 every target language, supporting aligned multilingual comparisons.

Translation procedure

Primary pass

• —Model: translategemma:27b (TranslateGemma 27B, Q4KM Ollama build)
• —Hardware: two NVIDIA L4 GPUs, one Ollama worker per GPU
• —Decoding: temperature 0.0
• —Context / generation cap: 4,096 / 4,096 tokens
• —Initial chunk limits: 4,500 source characters and 64 protected spans
• —Prompt style: TranslateGemma-specific
• —Accepted: 2,930/3,700 pairs

Fallback pass

• —Manifest: all 770 pairs not accepted by the primary pass
• —Model: gemma4:31b through Ollama
• —Hardware: two NVIDIA L4 GPUs, one Ollama worker per GPU
• —Context / generation cap: 8,192 / 8,192 tokens
• —Initial chunk limits: 4,500 source characters and 64 protected spans
• —Prompt style: generic chunked translation
• —Accepted: 495/770 pairs
• —Rejected: 208/770 pairs
• —Terminal errors: 67/770 pairs

Both passes used faithful_trace_translation. Long documents were translated in semantic chunks, reassembled, and accepted only after deterministic validation. Runtime values in translation_metrics are Ollama nanoseconds. Internal host values document the worker used and are not stable public endpoints.

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 structural markers;
6. 6.a plausible translated/source length ratio; and
7. 7.no high source-copy similarity indicating an untranslated output.

No accepted row crossed the likely-untranslated threshold. Final fallback rejection flags were unexpected script: 177, malformed source think tags: 37, expected script missing: 7, protected token loss: 1. A rejected row may carry more than one flag, so flag counts need not sum to the number of rejected rows. Representative terminal errors were protected-placeholder mismatches and generation-cap exhaustion. These gates detect structural and linguistic failures; they do not prove semantic equivalence or mathematical correctness of every intermediate statement.

Final per-language outcomes

Each language had 100 attempted source traces. Error denotes a pair that did not produce an accepted or deterministically rejected fallback result. Only accepted rows are published.

Dataset structure

The dataset has one train split containing conversational SFT rows. The translated user message contains the problem; the assistant message contains 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 3,425 published 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.

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 mathematical reasoning.
• —Source traces and their translations are model-generated, 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 detection.
• —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.
• —Coverage varies by language and missing pairs are not random; long traces and protected-token handling caused a substantial share of failures.
• —Two translation models and prompt styles are present; consumers can filter on translation_model when a homogeneous subset is required.

Validation still required before public release

1. 1.Stratified native-speaker review across language families and reasoning-length strata.
2. 2.Independent answer and solver-accuracy verification on a fixed multilingual sample.
3. 3.Semantic-equivalence checks focused on long, multi-chunk traces.
4. 4.Targeted review of low-coverage languages and fallback-model rows.
5. 5.PII and broader safety screening appropriate to the intended training use.

Reproducibility and versioning

v0.2-pilot is append-only relative to v0.1-pilot: all original accepted text is preserved and only newly accepted fallback rows are added. Stable IDs are derived from the source/language pair. Use translation_model and translation_metrics.prompt_style to distinguish stages. The machine-readable release manifest records exact counts and the Parquet checksum.

Citation

@misc{openeurollm_reasoning_traces_multilingual_pilot_2026,
  title        = {OpenEuroLLM Multilingual Mathematical Reasoning Traces: Two-Stage Pilot},
  author       = {{OpenEuroLLM contributors}},
  year         = {2026},
  howpublished = {Hugging Face dataset},
  note         = {Private v0.2-pilot staging release}
}

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