source_separation.md

JPN-Bench Source Separation

JPN-Bench source material must be kept separate from tokenizer training material. The crawler in this repo assigns whole source files to one lane:

• tokenizer training source files go to data/source_materials/tokenizer_training_sources.jsonl
• benchmark candidate source files go to benchmarks/jpn_bench/source_materials/benchmark_sources.jsonl

When content extraction is enabled, sentence/tree text follows the same split:

• tokenizer training text goes to data/source_materials/tokenizer_training_text.jsonl
• benchmark candidate text goes to benchmarks/jpn_bench/source_materials/benchmark_text.jsonl

The split is deterministic and based on source + source_file_id, not on individual sentences. This avoids placing neighboring sentences from the same article or source file into both training and benchmark lanes.

Configured Sources

The configured source list lives in configs/jpn_bench_source_materials.yaml. Kainoki per-file downloads are cached under data/raw/source_archives/kainoki/ so repeat crawls do not repeatedly hit the live CGI endpoint.

NPCMJ

The NPCMJ source uses the kana bracketed tree archive:

• https://www2.ninjal.ac.jp/npcmj/zip/npcmj_kana.zip

NPCMJ is a CC BY 4.0 research corpus. Keep the source file ID, archive URL, and citation metadata with any derived items.

Kainoki

The Kainoki source uses the live download index:

• https://oncoj.orinst.ox.ac.uk/cgi-bin/overview.sh?db=Kainoki&mode=download

Kainoki is CC BY 4.0 and is updated over time. Treat the index as live source material; record crawl dates in artifacts if publishing a benchmark release.

Commands

Build separated source-file manifests:

uv run jpn-tokenizer crawl-source-materials \
  --config configs/jpn_bench_source_materials.yaml

Download the NPCMJ archive and extract/fetch a small content sample:

uv run jpn-tokenizer crawl-source-materials \
  --config configs/jpn_bench_source_materials.yaml \
  --fetch-archives \
  --fetch-content \
  --max-content-files 20

For a full content crawl, omit --max-content-files. Do that deliberately: Kainoki requires hundreds of HTTP requests, so keep the request delay in the config and avoid rerunning full crawls unnecessarily.

Rules

• Never train a tokenizer on benchmarks/jpn_bench/source_materials/.
• Never promote benchmark items from source files assigned to the tokenizer training lane.
• When training with these sources, pass released benchmark JSONL files via --exclude-benchmark.
• Hidden benchmark sources should live outside the public repo and should use the same file-level split rule.
• For final benchmark items, prefer rewritten or hand-authored prompts inspired by the source material over raw copied sentences unless the license and benchmark design explicitly allow direct reuse.