All models
    MIST QG family · Question generation

    MIST-QG-1.0

    Only the questions a passage really supports, in 324 languages.

    MIST-QG-1.0 reads a passage and writes the questions it answers, in the passage’s own language or in English or French. It decides how many: a short passage gets one or two questions, a rich one more, so you only get questions the text can really answer. It is a 4B model trained on 324 languages for building retrieval training data.

    Also in MIST QG:MIST-QG-1.1
    Previous generation:mist-qg-1.5b
    MIST-QG-1.0 banner
    324
    Languages
    Auto
    The passage decides how many questions
    3
    Question languages: its own, English, French
    4B
    Parameters
    / Why it exists

    The problem it solves

    Retrieval models learn from pairs of a question and the passage that answers it, and a weak question makes a bad pair. Asking a model for a fixed number of questions invites padding: near-duplicates, or questions the passage only half answers.

    MIST-QG-1.0 lets the passage decide. It writes the questions the text supports and stops, with low-resource and African languages deliberately well covered.

    / Benchmarks

    The numbers

    Same language, all 238 test languages

    Same language, all 238 test languages

    47% of its questions are usable, as many as Qwen3.5-27B. Only Gemma 4 31B scores higher.

    Cross-lingual

    Cross-lingual

    English or French questions on passages in 153 other languages: 82% usable, the best of every model tested.

    / What you get

    Highlights

    As good as models seven times its size

    Averaged over 238 languages, 47% of its questions pass every check, as many as Qwen3.5-27B. Only Gemma 4 31B scores higher (51%).

    Strong where data is scarce

    31% usable in the rarest languages, against 15–25% for Qwen, gpt-oss and Llama models up to 70B.

    Cross-lingual

    82% of its English or French questions on passages in other languages are usable, the best of any model tested.

    Questions that find their passage

    Round-trip R@1 of 0.696, ahead of every other model tested.

    / Quick start

    Run it in a few lines

    $pip install vllm
    questions.py
    from vllm import LLM, SamplingParams
    
    PROMPT = ("Passage ({lang}):\n{passage}\n\n"
              "Questions in {qlang} that this passage answers, in your own words:\n")
    
    llm = LLM(model="olaverse/MIST-QG-1.0", max_model_len=4096)
    params = SamplingParams(max_tokens=256, temperature=0.3, top_p=0.9, stop=["\n\n"])
    
    prompts = [PROMPT.format(lang="Yoruba", qlang="Yoruba", passage=passage),    # same language
               PROMPT.format(lang="Yoruba", qlang="English", passage=passage)]   # cross-lingual
    for out in llm.generate(prompts, params):
        print([l[2:].strip() for l in out.outputs[0].text.split("\n") if l.startswith("- ")])
    / MIST QG family

    Which MIST QG model?

    Need a set number of questions, or questions into Spanish, Swahili, Hindi and 27 more languages? Use MIST-QG-1.1.

    MIST-QG-1.1 →MIST-QG-1.0THIS PAGE
    Parameters4B4B
    Passage languages324324
    How many questionsYou choose 1 to 5: exact every time cross-linguallyThe passage decides
    Question languagesOwn language, English, French, and 30 moreOwn language, English or French
    English or French → 30 languages59% usable, 99.3% in the right languageNot supported
    Between two of those 30 languages47% usableNot supported
    Cross-lingual questions per passageAbout 1.75About 1
    Questions that find their own passage (R@1)0.701, the best tested0.696
    Best forA set number of questions, in more languages and directionsLetting the passage decide how many
    On the Olaverse APIYesNo, self-host only
    / Where it shines

    Best for

    • Clean retrieval training pairs, where every question must be one the passage really answers.
    • Same-language questions over passages in 324 languages, low-resource and African languages included.
    • English or French questions on passages in any of those languages, for cross-lingual search.
    / Built with
    Qwen3.5-4B-Base vLLM Transformers Apache-2.0