chiboard-1-sft

A large-scale supervised fine-tuning dataset for Chinese IME (pinyin-to-Hanzi) text conversion. Given a pinyin (or mixed pinyin/Chinese) active_input together with an already-decoded committed_context, the model must predict the Chinese target text. It is designed to train and evaluate models that power a rolling decode-window IME, where text is committed in chunks and the active region is retyped/edited as the user goes.

• Total rows: 15,969,584 (train 15,330,235 · test 320,091 · dev 319,258)
• Built: 2026-07-03 (seed 20260703)
• Format: Parquet (8 train shards + 1 dev + 1 test), plus JSONL inspection samples and a QA report
• Language: Mandarin Chinese (zh) with romanized pinyin inputs
• All QA gates passed (see qa_report.md)

Task formulation

For each row, the model receives:

• committed_context — already-finalized Chinese text to the left of the cursor (may be empty).
• active_input — the current, not-yet-committed buffer: pinyin, mixed pinyin+Chinese, mid-syllable fragments, or already-Chinese (depending on variant).

…and must produce:

• target — the correct Chinese realization of active_input, consistent with committed_context.

Boundary semantics: the committed/active boundary follows rolling decode-window retirement — there are no explicit user-commit events. Cut preference when retiring the active region: sentence-final punctuation > clause comma > word boundary. This matches a deployment where the IME commits chunks of 40–100 characters; 31.82% of rows have len(active_input) in the 40–100 range.

Row schema

field	type	description
split	string	train / test / dev
committed_context	string	Finalized Chinese text left of the cursor (empty for 58.8% of rows)
active_input	string	The active buffer to convert (pinyin / mixed / fragments / Chinese)
target	string	Ground-truth Chinese output
source	string	Source corpus id (see Source data)
source_document_id	string	Stable document id within the source; used for split-leakage control
variant	string	Input transformation variant (see Variants)
noise	object	{types: [...], edit_count: int} — injected typos / mid-text edits
hard_ambiguity_terms	sequence[string]	Hard-ambiguity pinyin terms present (e.g. shi, zai, de_di_de)
length_chars	int64	Character length of the row
id	string	Unique row id, e.g. train_14726950

Variants

Input variant distribution (target vs achieved; all gates passed):

variant	share	rows	noise rate
joined_pinyin	0.3511	5,606,255	0.0
mixed_chinese_pinyin	0.2807	4,483,376	0.0997
noisy_pinyin	0.1204	1,921,974	1.0
mid_syllable	0.0776	1,239,560	0.1027
punctuation_removed	0.0700	1,118,436	0.0
ascii_punctuation	0.0500	798,910	0.0
syllable_spaced_pinyin	0.0301	480,635	0.0
already_chinese	0.0201	320,438	0.0

Length distribution (chars)

bucket	share	rows
2–8	0.1902	3,038,075
9–30	0.3354	5,356,188
31–120	0.3013	4,812,020
121–500	0.1259	2,009,839
501–1500	0.0379	605,571
1501–3500	0.0093	147,891

Coverage and ambiguity

• Top-50,000 headwords: all 50,000 meet the ≥200 train-rows quota.
• Top-10,000 headwords: all 10,000 meet the ≥1,000 train-rows quota.
• Hard-ambiguity quotas (all PASS, min 30,000 rows each):

term	rows
de_di_de	8,701,822
shi	6,084,655
zai	4,087,093
yao	2,532,302
ta	1,792,026
zuo	2,207,748
jiao	1,252,547
xian	1,185,995
shijian	589,666
qishi	207,641
changshi	47,964
hangxing	30,009

5,411 contrastive near-pair groups were mined to stress disambiguation.

Realism properties

• committed_context non-empty: 41.22% (gate min 30%)
• Intra-chunk retirement: 43.87% of context rows retire inside the chunk (gate min 25%)
• Active region in 40–100 chars: 31.82% (gate min 25%)
• Mixed-mode rows: tail 5,704,045 vs mid-text edit 1,298,617 (edit share 18.54%, gate min 10%)
• Punctuation style over pinyin rows: ascii 65.4% / cjk+mixed+none 34.6%
• ü→v vs ü→u policy: u-form 10.01% of v-eligible syllables (bounds 5–15%)
• No scattered 汉字 inside typed pinyin regions (gate PASS)
• No split leakage: 0 document ids appear in more than one split
• 65,675 exact prompt+target duplicates dropped globally

Source data

source	rows	license	url
lccc_large	7,656,904	mit	https://huggingface.co/datasets/thu-coai/lccc
chinesewebtext2_hq	6,517,813	apache-2.0	https://huggingface.co/datasets/Morton-Li/ChineseWebText2.0-HighQuality
dureader_retrieval_corpus	1,401,928	apache-2.0	https://huggingface.co/datasets/zyznull/dureader-retrieval-corpus
clapai_sentiment_zh	203,522	apache-2.0	https://huggingface.co/datasets/clapAI/MultiLingualSentiment
somebreeze_news	102,807	apache-2.0	https://huggingface.co/datasets/somebreeze/Chinese-news-summery
feilongfl_news	23,327	apache-2.0	https://huggingface.co/datasets/feilongfl/ChineseNewsSummary
dureader_retrieval_queries	44,370	apache-2.0	https://huggingface.co/datasets/zyznull/dureader-retrieval-corpus
cmrc2018	18,913	cc-by-sa-4.0	https://huggingface.co/datasets/hfl/cmrc2018

Rejected sources (and why) are listed in metadata.json under rejected_sources (e.g. lccc_base_train is a subset of lccc_large; lucaswychan_dureader_hard_neg overlaps dureader_retrieval_corpus).

Files

• train-00000.parquet … train-00007.parquet — train split (8 shards)
• dev.parquet — dev split
• test.parquet — test split
• metadata.json — full build metadata, distributions, gate results, source manifest
• qa_report.md — human-readable QA report (all gates PASS)
• sample_500.jsonl — 500-row inspection sample
• polyphone_sample_500.jsonl — 500 rows containing polyphone characters for manual verification

Notes

• A held-out reserve partition of 8,970,859 rows exists locally (report-only, no gates) for future expansion; it is not included in this repo.
• coverage-fill upsampled rows: 0. ambiguity-fill rows: 16,287 (intentional reweighting).
• Build elapsed time: ~699 s.

Citation

If you use this dataset, please cite the underlying sources listed in the table above.