olaverse/qg-passages-multi

qg-passages-multi

The training set behind `olaverse/mist-qg-1.5b` — passages paired with search-style questions they directly answer, across 25 languages.

Dataset Summary

~50,000 passages, each with 2–3 natural-language questions that the passage answers. Built by distilling a strong teacher model over Aya Collection passages, then filtering with round-trip retrieval verification so every kept question is a genuine, answerable positive for its passage.

Data Fields

Data Instance

{
  "passage": "Tides are caused by the gravitational pull of the moon and, to a lesser extent, the sun, acting on Earth's oceans.",
  "questions": ["What causes ocean tides?", "Does the sun affect tides?", "Which has a bigger effect on tides, the moon or the sun?"],
  "language": "eng",
  "teacher": "Qwen/Qwen2.5-32B-Instruct"
}

Languages

25 languages, ~2,000 source passages each: English, French, German, Spanish, Portuguese, Italian, Dutch, Russian, Polish, Turkish, Vietnamese, Indonesian, Hindi, Japanese, Korean, Yoruba, Igbo, Hausa, Swahili, Amharic, Zulu, Xhosa, Shona, Somali, Afrikaans.

This dataset is multilingual, not cross-lingual — every question was generated from a single passage in a single language. There is no signal here for cross-lingual (query in one language, passage in another) retrieval.

Dataset Creation

1. 1.Source passages: streamed from `CohereLabs/aya_collection_language_split` (Apache-2.0).
2. 2.Question generation: Qwen/Qwen2.5-32B-Instruct (Apache-2.0) writes 2–3 questions per passage, prompted to vary type (factual / yes-no / comparison) and avoid meta-references like "according to the passage."
3. 3.Verification: each question is embedded with Qwen/Qwen3-Embedding-0.6B and checked via round-trip retrieval — kept only if it retrieves its own source passage out of [source + 100 distractors] at rank 1. Keep-rates ranged 0.93–1.00 for most languages; lower for a few lower-resource ones.

Known Limitations

A subset of passages sourced from Aya's instruction-tuning-derived subsets retain leftover template artifacts from that construction (e.g., a trailing "answer the following questions: 1. ... 2. ..." block appended to otherwise-clean prose). This affects roughly 10–25% of passages for most languages, and considerably more for Korean and Amharic specifically (~94% and ~93.6% respectively).

A cleaned variant (regex-stripped template tails) was built and evaluated during model development. The model trained on the cleaned variant scored lower on held-out evaluation than the model trained on this uncleaned version — the stripping heuristic appears to have removed real content along with template junk for the most heavily-affected languages. This dataset intentionally ships the original, uncleaned version as a result. See the `mist-qg-1.5b` model card for the evaluation this decision was based on.

Licensing

Apache-2.0. Source passages from Aya Collection (Apache-2.0); questions generated by Qwen2.5-32B-Instruct (Apache-2.0).

Citation

@misc{qg-passages-multi,
  title  = {qg-passages-multi},
  author = {Olaverse},
  year   = {2026},
  url    = {https://huggingface.co/datasets/olaverse/qg-passages-multi}
}