chiboard-1-sft

A large-scale supervised fine-tuning dataset for Chinese IME (pinyin-to-Hanzi) text conversion. Given a committed Chinese context plus a pinyin/mixed active buffer, 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: 59,293,523 (train 56,918,955 · test 1,184,732 · dev 1,189,836)
• Built: 2026-07-08 (seed 20260703; rows globally shuffled)
• Format: Parquet — 30 train shards (412 MB each) + 1 dev + 1 test (13 GB total), plus JSONL inspection samples and a QA report
• Tokenizer: Loss-token metrics measured with LiquidAI/LFM2.5-350M-Base
• Language: Mandarin Chinese (zh) with romanized pinyin inputs

Scale

Rows	Total tokens	Loss tokens	Avg loss/row	Short token share (2–30 chars)
59,293,523	4.91B	1.32B	22.2	49.2%

Splits

split	rows	share
train	56,918,955	96.0%
dev	1,189,836	2.0%
test	1,184,732	2.0%

Prompt format

Rows are formatted into the following template before being fed to the model:

<|startoftext|>{committed_context}<|reserved_6|>{raw_pinyin}<|reserved_7|>{display}<|reserved_8|>{target}<|im_end|>

The model is trained to produce {target} (loss is taken on the target span). Reserved tokens separate committed context, raw pinyin buffer, display buffer (mixed pinyin+Chinese as currently shown), and the gold Chinese target.

Task formulation

Boundary semantics: the committed/active boundary follows rolling decode-window retirement — there are no explicit user-commit events. Retirement cuts are sentence-final only (context_sentence_boundary_share = 1.0). Deployment window sizes are 40–100 chars; 31.79% of rows have len(display) in that range.

For each row the model receives:

• committed_context — finalized Chinese text left of the cursor (non-empty for 7.1% of rows).
• raw_pinyin — the pure pinyin buffer to convert.
• display — the current on-screen mixed pinyin+Chinese rendering of the active region.

…and must produce target — the correct Chinese realization.

Row schema

field	type	description
split	string	train / test / dev
committed_context	string	Finalized Chinese text left of the cursor
raw_pinyin	string	Pure pinyin input buffer
display	string	Mixed pinyin+Chinese display string for the active region
target	string	Ground-truth Chinese output
source_text	string	Original Chinese source text the row was derived from
source_pinyin	string	Pinyin of the source text
source	string	Source corpus id (see Source data)
source_document_id	string	Stable document id; 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
length_chars	int64	Character length of the row
id	string	Unique row id

Variants

variant	share	rows	noise rate
joined_pinyin	0.3798	22,518,080	0.0
mixed_chinese_pinyin	0.2923	17,331,916	0.10
noisy_pinyin	0.1120	6,642,954	1.0
mid_syllable	0.0643	3,812,774	0.1127
punctuation_removed	0.0627	3,718,125	0.0
ascii_punctuation	0.0445	2,637,085	0.0
syllable_spaced_pinyin	0.0267	1,580,306	0.0
already_chinese	0.0177	1,052,283	0.0

Length distribution (chars)

bucket	share	rows
2–8	0.2621	15,538,393
9–30	0.6054	35,897,780
31–120	0.1069	6,340,406
121–500	0.0215	1,275,339
501–1500	0.0035	206,769
1501–3500	0.0006	34,836

Active region length (display, chars)

Deployment window sizes are 40–100 chars.

length (chars)	rows	share
1–9	3,208,702	5.4%
10–19	15,211,958	25.7%
20–39	17,213,036	29.0%
40–69	13,878,187	23.4%
70–100	4,972,345	8.4%
101–160	2,027,128	3.4%
161–300	1,891,156	3.2%
301+	891,011	1.5%

Hard-ambiguity coverage

Rows containing each hard-ambiguity pinyin term (minimum 30,000 per term):

term	rows
de_di_de	23,244,998
shi	16,468,963
zai	7,985,426
yao	5,662,289
ta	3,420,909
zuo	3,154,146
jiao	1,700,574
xian	1,542,407
shijian	816,010
qishi	442,134
changshi	60,112
hangxing	40,861

8,194 contrastive near-pair groups mined for disambiguation stress-testing.

Punctuation styles (pinyin rows)

style	rows	share
ascii	22,813,334	39.1%
na (no punct)	22,634,009	38.8%
cjk	4,236,244	7.3%
mixed	4,839,528	8.3%
none (removed)	3,718,125	6.4%

Dataset properties

Coverage

• Top-50,000 headwords: 50,000/50,000 meet ≥200 train-row quota.
• Top-10,000 headwords: 10,000/10,000 meet ≥1,000 train-row quota.

Context & boundaries

• Rows with committed_context: 4,195,414 (7.08%).
• Context sentence/turn boundary alignment: 100%.
• Context types: chunk 3,052,869 · intra-sentence 1,142,545.
• Retirement cuts: 1,907,770 (all sentence-final).
• Intra-chunk retirement share: 27.23%.
• Conversational single-turn share: 99.47%.

Mixed-mode & noise

• Mixed tail rows: 25,381,550 · mid-text edit rows: 5,259,785 (edit share 17.17%).
• Noisy rows (any injected edit): 8,805,910 (14.9%).
• ü→v vs ü→u: u-form 567,130 · v-form 5,101,515 (u-share 10.0%).

Slicing & multiplicity

• Slice rows: 7,065,368 (cut from long formal entries at sentence-preferred boundaries).
• Up to 4 distinct rows per source pool entry (conversational-first assignment).
• 855,727 exact prompt+target duplicates dropped (0.9% of pre-dedup pool).

Register (target char mass by source)

Conversational sources carry 44.2% of target char mass.

source	rows	target chars	char share
chinesewebtext2_hq	7,546,105	468.1M	33.8%
lccc_large	32,585,973	429.4M	31.0%
dureader_retrieval_corpus	4,332,104	260.5M	18.8%
personal_dialog	13,526,191	167.8M	12.1%
clapai_sentiment_zh	581,719	18.5M	1.3%
all others	1,235,431	41.2M	3.0%

Split integrity

• Doc-keyed splits: 0 document ids leak across train/dev/test.
• Rows globally shuffled within each split.

Source data

source	rows	license	url
lccc_large	32,585,973	mit	https://huggingface.co/datasets/thu-coai/lccc
personal_dialog	13,526,191	other (research-only)	https://huggingface.co/datasets/silver/personal_dialog
chinesewebtext2_hq	7,546,105	apache-2.0	https://huggingface.co/datasets/Morton-Li/ChineseWebText2.0-HighQuality
dureader_retrieval_corpus	4,332,104	apache-2.0	https://huggingface.co/datasets/zyznull/dureader-retrieval-corpus
clapai_sentiment_zh	581,719	apache-2.0	https://huggingface.co/datasets/clapAI/MultiLingualSentiment
somebreeze_news	286,942	apache-2.0	https://huggingface.co/datasets/somebreeze/Chinese-news-summery
lccc_dialogue	154,592	mit (derived)	https://huggingface.co/datasets/thu-coai/lccc
dureader_retrieval_queries	85,014	apache-2.0	https://huggingface.co/datasets/zyznull/dureader-retrieval-corpus
pd_dialogue	88,827	other (research-only)	https://huggingface.co/datasets/silver/personal_dialog
feilongfl_news	54,556	apache-2.0	https://huggingface.co/datasets/feilongfl/ChineseNewsSummary
cmrc2018	51,500	cc-by-sa-4.0	https://huggingface.co/datasets/hfl/cmrc2018

Files

• train-00000.parquet … train-00029.parquet — train split (30 shards, ~412 MB each)
• dev.parquet — dev split (~258 MB)
• test.parquet — test split (~257 MB)
• metadata.json — build metadata, distributions, and source manifest
• qa_report.md — QA report
• sample_500.jsonl — 500-row inspection sample
• polyphone_sample_500.jsonl — 500 rows with polyphone characters for manual verification

Citation

If you use this dataset, please cite the underlying sources listed above, with particular attention to PersonalDialog (Zheng et al., 2019) for the research-only conversational component.