Infomaniak-AI/speculators-multilingual-en-fr-de-it-es

Speculators Multilingual SFT Dataset (en/fr/de/it/es)

A multilingual instruction-following dataset in ShareGPT format, built to train draft models for speculative decoding across English, French, German, Italian and Spanish.

Summary

An English instruction-tuning corpus with part of it kept in English and the rest machine-translated into French, German, Italian and Spanish using `tencent/Hunyuan-MT-7B`. Provided as a single mixed-language, ShareGPT-formatted dataset, for training draft models used in speculative decoding.

Dataset Structure

Each example is a single-turn conversation:

{
  "id": "sample_XXXXXX",
  "language": "fr",
  "conversations": [
    {"from": "human", "value": "5020 moins un nombre divisé par 20,08 équivaut à 4970. Quel est ce nombre ?"},
    {"from": "gpt", "value": "Pour trouver ce nombre, nous pouvons établir l’équation suivante :\n\n5020 – (x / 20,08) = 4970\n\n..."}
  ]
}

Fields:

• —id: source row identifier (informational only)
• —language: one of en, fr, de, it, es
• —conversations: a two-turn human/gpt exchange, ShareGPT-style

Splits (languages are mixed within each file, in roughly equal proportion):

Source Data & Methodology

• —Base corpus: `mgoin/open-perfectblend-glm5.2-regen`, a ShareGPT-style instruction dataset. Only the first human/assistant exchange of each conversation is used here.
• —Cleaning: any ...</think> reasoning prefix is stripped from responses, keeping only the final answer.
• —Translation: French, German, Italian and Spanish slices are produced with `tencent/Hunyuan-MT-7B`, served via vLLM, using the model's plain recommended prompt template ("Translate the following segment into <language>, without additional explanation.") and its recommended sampling settings (temperature 0.7, topp 0.6, topk 20, repetition penalty 1.05).
• —Length cap: both the prompt and the response are limited to ~1400 tokens (measured with Hunyuan-MT-7B's own tokenizer) — the range where translation quality was empirically verified to be reliable.
• —Quality filter: each translation is checked against its original text via a translated/original word-count ratio; translations that come out too short (a proxy for truncation or summarization) are discarded rather than kept.

Prompt length (characters, across all languages): median ≈ 274, 10th percentile ≈ 86, 90th percentile ≈ 1,321.

Intended Use

Training/distilling draft models for speculative decoding.

Limitations

• —Translations are machine-generated (no human review); quality is generally strong within the length range used, but automatic translation always carries some risk of subtle inaccuracies.

License

This dataset is derived from `mgoin/open-perfectblend-glm5.2-regen`; please refer to that dataset's license terms, which apply here as well.

Citation

If you use this dataset, please also credit:

• —The source dataset: mgoin/open-perfectblend-glm5.2-regen
• —The translation model: tencent/Hunyuan-MT-7B