jiviteshjn/mc4-zh-idiom-cpt

mC4 zh — Idiom-Tagged Continued-Pretraining Corpus

A 9.6M-document Chinese corpus for continued pretraining on cultural knowledge in figurative language. Each document is natural web text (from the C4/mC4 zh subset) containing at least one culturally meaningful chengyu, with an appended knowledge block that lists every matched idiom together with its figurative meaning(s) and classical source citation.

Built 2026-07-16 as Stage 1 (continue-pretraining data) of the CultureInFigurativeLanguage project.

Key numbers

Documents	9,609,311
Tokens (Qwen3.5 tokenizer, measured on a sample)	~14.4B
Compressed size	~27 GB (1,024 gzip files)
Idiom annotations	32,048,011
Distinct idioms covered	26,727 (of a 27,297 vetted list)
Per-idiom document cap	10,000 (rare-idiom-first selection)
Source corpus	allenai/c4, zh train split (54.5M docs)

Document format

data/tagged_*.json.gz — one JSON object per line:

{
  "text": "<original document text>\n\n【成语注释】\n<idiom>：<figurative meaning(s)>（出处/字面：<classical source>）\n...",
  "url": "...",
  "timestamp": "...",
  "source": "mc4-zh",
  "shard": 0,
  "line": 123,
  "matched_idioms": ["万事大吉", "..."],
  "original_text_chars": 883
}

text is the training-ready field (document + knowledge block). The block creates explicit co-occurrence between each idiom and its figurative meaning; original_text_chars lets you recover the untagged document if needed.

How it was built

1. Idiom vetting — a merged 31K-entry Chinese chengyu dataset (xinhua, chengyu_data, FuxiBench, IdiomKB) was filtered by an LLM judge (gpt-5.4-mini) on cultural/figurative content, not commonness: entries kept if they carry figurative meaning beyond the literal reading, derive from a classical allusion (典故), or reflect culturally significant imagery; politeness formulas, grammatical connectives, and compositional phrases removed. Result: 27,297 idioms. Per-entry verdicts with reasons: idioms/filter_decisions_zh.jsonl; the vetted list with meanings: idioms/idioms_merged_llm_formatted_figurative_only.jsonl.
2. Candidate extraction — the full C4 zh train split (54.5M docs) was scanned with Aho-Corasick multi-pattern matching against the vetted list. Quality gates: 150–100,000 chars, and a word-list gate dropping documents matching >20 distinct idioms (dictionary/word-list pages). 15.9M candidates.
3. Deduplication — MinHash-LSH near-dedup (128 perms over UTF-8 byte 16-gram shingles, 13 bands × 9 rows ≈ 0.75 Jaccard threshold, drops verified against the cluster representative): 591K near-duplicates removed.
4. Frequency-capped selection (rare-idiom-first) — idioms processed in ascending candidate count, each topping up to 10,000 documents from its own shuffled candidates; a selected document credits all idioms it contains, so common idioms fill via co-occurrence and never crowd out rare, culturally specific ones. Kept counts may exceed the cap through co-occurrence credit. Per-idiom kept counts: stats/kept_idiom_counts_zh.json.
5. Metadata tagging — each selected document received the 【成语注释】 knowledge block: up to 2 deduplicated figurative meanings per idiom plus the first literal meaning / classical source citation.

Pipeline summaries for every stage are in stats/.

Usage

from datasets import load_dataset
ds = load_dataset("jiviteshjn/mc4-zh-idiom-cpt", data_files="data/*.json.gz", split="train")

Or download for local training:

hf download jiviteshjn/mc4-zh-idiom-cpt --repo-type dataset --local-dir mc4_zh_idiom_cpt

Documents average ~1,500 tokens; sequence packing is recommended for training.

Provenance & license

Derived from allenai/c4 (ODC-BY, Common Crawl web text — inherits its content caveats, including residual spam/SEO prose). Idiom meanings merged from public chengyu dictionaries. Intended for research on cultural alignment of language models.