lukfre/eurobench-parallel

EuroBench Parallel

This repo contains several multilingual parallel benchmarks covering European languages ready to be used with lm-eval-harness.

Collection

As of October, 2026, it includes

• MMLU-mix, our own build, cleaned and refined from:

  • cais/mmlu
  • openai/mmmlu
  • CohereLabs/Global-MMLU
  • edinburgh-dawg/mmlu-redux-2
  • collections/INSAIT-Institute/multilingual-benchmarks
  • Eurolingua/mmlux
  • EC-DGT-AI/EU-MMLU (Croatian and Irish only)

  Croatian (mmlu_hr, 642 items) and Irish (mmlu_ga, 287 items) are partial: they have only the items their source covers, so their scores are not comparable with the other languages' (5,315 items) and should be reported separately.

• BELEBELE, copied from facebook/belebele

• ARC-Easy / ARC_Challenge, copied from allenai/ai2_arc + Eurolingua/arcx

• PolyMath, copied from Qwen/PolyMath

Schema

Every multiple-choice benchmark gets the same schema, so one set of lm-eval templates covers all:

    <data>
    id               item key shared by all languages (mmlu "subject/idx", belebele "link#qnum", ARC id)
    question         question text (the passage is separate, in `passage`, for belebele)
    choices          option texts, already in the NEW order
    answer           index of the correct option in `choices` (NEW order)
    answer_original  index of the correct option in the source order
    perm             perm[new_position] = source position, so choices[i] == source_choices[perm[i]]
    shuffled         False when the item kept its source order (see "Pinned options")

    <metadata>
    lang
    translation_tier
    source_dataset

    <custom>
    + mmlu: subject, subject_category, idx
    + belebele: passage

Shuffle rules

Some source distributions are skewed: for example, MMLU-Redux-EU has D 27.6%, so "always D" would beat chance.

• One permutation per item, decided on the English item and applied to every language, so the files stay line-parallel: the i-th row has the same correct letter everywhere.
• The correct option always moves to a different position than in the source.
• Its new position is chosen greedily to balance the answer letters within each split. Distractors fill the remaining positions at random.
• Pinned options: an option that refers to the others ("All of the above", "Both A and B", "Neither of these") keeps its source position; the rest are shuffled around it. If the correct answer itself is pinned it cannot move, so the item keeps its source order (shuffled=False).
• Deterministic: items are processed in an order fixed by a hash of (seed, id).
• Few-shot splits drop items with an empty option in any language (in every language); test splits are kept whole.

Evaluation settings

We are collecting both generative- and likelikhood-based benchmarks, to cover the evaluation of the base and fine-tuned models.

PolyMath is free-form maths (no options): copied as is, one test split with a level column.