chinese-babylm-org/chinesebabylm-2026-leaderboard

ChineseBabyLM 2026 Leaderboard

Leaderboard for the ChineseBabyLM 2026 Challenge.

Evaluation Tracks

• —NLU Track: ZhoBLiMP (zero-shot) + CLUE benchmarks (AFQMC, OCNLI, TNEWS, CLUEWSC2020)
• —Cognitive Modeling Track: Brain-aligned fMRI evaluation (word_fmri, fmri)
• —HANZI Track: Character structure and phonology (hanzistructure, hanzipinyin)

Submission Format

Upload a JSON file with pre-computed scores:

{
    "zhoblimp": {"accuracy": 0.75},
    "afqmc": {"accuracy": 0.70},
    "ocnli": {"accuracy": 0.65},
    "tnews": {"accuracy": 0.60},
    "cluewsc2020": {"accuracy": 0.55},
    "word_fmri": {"mean": 0.30},
    "fmri": {"mean": 0.25},
    "hanzi_structure": {"accuracy": 0.60},
    "hanzi_pinyin": {"accuracy": 0.55}
}

Use the chinese-babylm-eval-pipeline to run evaluations and collate results.

Configuration

• —src/envs.py — HuggingFace repository paths
• —src/about.py — Task definitions and leaderboard text
• —src/display/utils.py — Leaderboard column configuration