CohereLabs/m-ArenaHard-v2.1

Dataset Card for m-ArenaHard-v2.1

The m-ArenaHard-v2.1 dataset is a multilingual LLM evaluation set built from the LMarena arena-hard-auto-v2.0 prompts used in m-ArenaHard-v2.0. It keeps the public v2.0 row schema while expanding coverage to the 67 raw translation files produced for the Tiny Aya evaluation work.

The dataset includes 67 languages:

am, ar, bg, bn, ca, cs, cy, da, de, el, en, es, et, eu, fa, fi, fr, ga, gl, gu, ha, he, hi, hr, hu, id, ig, it, ja, jv, km, ko, lo, lt, lv, mg, mr, ms, mt, my, ne, nl, no, pa, pl, pt, ro, ru, sk, sl, sn, sr, sv, sw, ta, te, th, tl, tr, uk, ur, vi, wo, xh, yo, zh, zu.

Each config contains a test split with 498 prompts.

Source and Quality Note

This dataset is derived from the Arena-Hard v2.0 prompts released by LMarena. In this multilingual version, we translated those prompts into the listed languages and applied a set of targeted fixes for issues we observed during processing, including code-preservation problems, corrupted or mixed-language English source prompts, and translation artifacts.

We also acknowledge that the upstream prompt set is not uniformly clean. Some quality issues originate in the original English data and may not be fully eliminated by downstream translation and repair. The fixes in this release are intended to make the benchmark more usable and consistent for multilingual research while preserving row alignment with the original Arena-Hard v2.0 source; users should still treat individual prompts as benchmark data with known provenance and limitations.

What Changed From v2.0

Compared with m-ArenaHard-v2.0, this version focuses on broader language coverage and translation-quality fixes while preserving the same public row schema.

Key changes:

• Expanded to 67 language configs, including all raw translation files available for this release.
• Repaired code-heavy and data-heavy prompts where machine translation had changed executable syntax, identifiers, SQL, JSON keys, formulas, stack traces, tables, or code fences.
• Added code/data-aware validation so natural-language prompt text is translated while task input such as source code, raw data, API names, filenames, and quoted literals remains intact.
• Audited the English source prompts inherited from v2.0. A small number of English rows were non-English, mixed-language, or corrupted, including Vietnamese instruction prefixes in English math/statistics prompts; v2.1 uses canonical English source overrides for those rows before rebuilding aligned multilingual prompts.
• Reconstructed one corrupted English math prompt (question_id: f3176f24b8364008, en_index=285) as a clean oil-tank queue problem and regenerated it across all 67 languages, preserving row alignment instead of removing the sample.

The final local validation for this release candidate checked that all 67 language configs load with 498 rows, the public schema is unchanged, accepted repairs have no deterministic structural failures, and the targeted English-source fixes no longer contain the known stale or corrupted text.

Dataset Structure

Each row has the same public fields as m-ArenaHard-v2.0:

{
    "question_id": "2edbb5f36f5b42be",
    "category": "hard_prompt",
    "subcategory": "coding",
    "prompt": "Write me a zig program...",
    "language": "en",
}

Dataset Fields

• question_id: a stable unique ID for the English source prompt. The same question_id is used across languages.
• category: prompt category from the original dataset.
• subcategory: finer-grained prompt category from the original dataset.
• prompt: translated prompt text.
• language: language config code.

Load With Datasets

from datasets import load_dataset

dataset = load_dataset("CohereLabs/m-ArenaHard-v2.1", "es", split="test")

To load every language:

from datasets import concatenate_datasets, get_dataset_config_names, load_dataset

moniker = "CohereLabs/m-ArenaHard-v2.1"
dataset = concatenate_datasets(
    [load_dataset(moniker, config, split="test") for config in get_dataset_config_names(moniker)]
)