JPN-Bench

JPN-Bench is a Japanese literacy benchmark for tokenizer evaluation and future Japanese LLM evaluation. This public release contains a small curated tokenizer-literacy dev set plus benchmark-lane source material manifests kept separate from tokenizer training material.

This dataset is grouped with the KotodamaLM tokenizer work in the Hugging Face collection "KotodamaLM Japanese Language Infrastructure".

Files

• data/literacy_items.jsonl: 60 tokenizer-literacy items balanced across n5, n4, and above_n4.
• data/seed_items.jsonl: 3 schema smoke-test items.
• data/source_materials/benchmark_sources.jsonl: source-file manifest for benchmark candidate material.
• data/source_materials/benchmark_text.jsonl: extracted benchmark-lane text.
• spec.md: benchmark schema and scoring intent.
• scale_plan.md: roadmap toward Lite, Verified, Full, SWE-sized, and Hidden tiers.
• scale_targets.json: machine-readable target counts.
• source_separation.md: rules that keep benchmark and tokenizer-training material apart.

Loading

from datasets import load_dataset

literacy = load_dataset(
    "MarcoDotIO/jpn-bench",
    "literacy",
    split="dev",
)

For local or explicit loading:

from datasets import load_dataset

dataset = load_dataset(
    "json",
    data_files={
        "literacy": "data/literacy_items.jsonl",
        "seed": "data/seed_items.jsonl",
        "benchmark_sources": "data/source_materials/benchmark_sources.jsonl",
        "benchmark_text": "data/source_materials/benchmark_text.jsonl",
    },
)

Separation Policy

Do not train tokenizers or LLMs on the public JPN-Bench files if you intend to report JPN-Bench scores. The paired KotodamaLM tokenizer was trained from the training lane and then aggressively filtered to remove any released JPN-Bench target surfaces.

The source split is file-level and deterministic. Whole source files are assigned to either the tokenizer-training lane or benchmark lane, preventing neighboring sentences from the same source file from landing in both places.

Current Scope

The public literacy set is intentionally small:

• 60 dev items
• 209 tokenization target surfaces
• 20 items each for n5, n4, and above_n4

The scale plan targets:

• 300-item Lite
• 500-item Verified
• 2,500-item Full
• optional 2,294-item SWE-sized Full slice
• private Hidden tier for saturation checks

Sources And Attribution

Benchmark-lane source material is derived from CC BY 4.0 Japanese corpora:

• NINJAL Parsed Corpus of Modern Japanese (NPCMJ)
• Kainoki / Open National Corpus of Japanese related source materials

See the JSONL source manifests for per-source URL, license, and citation fields.

Citation

@dataset{jpn_bench_2026,
  title = {JPN-Bench},
  author = {MarcoDotIO},
  year = {2026},
  publisher = {Hugging Face},
  license = {CC-BY-4.0}
}