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: 56,810,186 (train 54,535,518 · test 1,135,059 · dev 1,139,609)
• Built: 2026-07-09 (seed 20260703; rows globally shuffled)
• Format: Parquet — 30 train shards (349 MB each) + 1 dev + 1 test (11 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

Serving contract: empty display means display == raw_pinyin

For every pure-pinyin row (variants joined_pinyin, noisy_pinyin, punctuation_removed, ascii_punctuation, syllable_spaced_pinyin, and any other row whose on-screen state is identical to the raw buffer) the display column — and therefore the <|reserved_7|> slot in the serialized prompt — is empty. 79.7% of rows use this convention.

Consumers (training script, demo app, runtime) must send an empty <|reserved_7|> slot whenever the on-screen text equals the raw pinyin buffer. A model trained on this build sees populated display only when it differs from raw_pinyin (mixed 汉字+pinyin, mid-syllable with 汉字 prefix, already-Chinese, and noisy rows whose screen state diverged). Sending the duplicated display string is out-of-distribution.

The reserved tokens are always emitted, so the frame layout is unchanged:

<|startoftext|>…<|reserved_6|>zhege<|reserved_7|><|reserved_8|>这个<|im_end|>

See also metadata.json → serving_contract.

Scale

Rows	Prompt tokens	Loss tokens	Loss fraction	Short token share (2–30 chars)	Empty display share
56,810,186	3.89B	1.29B	34.6%	49.4%	79.7%

Loss-token and prompt-token counts are for the SFT partition; measured with LiquidAI/LFM2.5-350M-Base. Empty-display rows save ~1.74B prompt tokens vs duplicating raw_pinyin in the display slot.

Splits

split	rows	share
train	54,535,518	96.0%
dev	1,139,609	2.0%
test	1,135,059	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|>

• committed_context — trailing (≤64 chars) fully-converted 汉字 retired by the rolling decode window; empty for most short rows.
• raw_pinyin — the raw typed bytes of the active region (source of truth).
• display — what is on screen: provisional 汉字 + unconverted tail. Empty when identical to raw_pinyin (see serving contract above).
• target — gold conversion of the active region (loss is computed on target + <|im_end|> only).

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; 30.64% of rows have len(display) in that range.

For each row the model receives committed_context, raw_pinyin, and display (possibly empty), and must produce target.

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 on-screen rendering; empty when equal to raw_pinyin
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	0.5487	31,172,520	0.0
mixed_chinese_pinyin	0.1551	8,811,285	0.0994
noisy_pinyin	0.0771	4,380,150	1.0
mid_syllable	0.0681	3,869,283	0.1141
punctuation_removed	0.0613	3,480,239	0.0
ascii_punctuation	0.0431	2,449,231	0.0
syllable_spaced_pinyin	0.0280	1,589,316	0.0
already_chinese	0.0186	1,058,162	0.0

Length distribution (chars)

bucket	share	rows
2–8	0.2560	14,541,379
9–30	0.6058	34,417,375
31–120	0.1115	6,334,422
121–500	0.0225	1,275,407
501–1500	0.0036	206,767
1501–3500	0.0006	34,836

Active region length (display, chars)

length (chars)	rows	share
1–9	2,685,513	4.7%
10–19	12,245,986	21.6%
20–39	18,781,726	33.1%
40–69	12,922,293	22.7%
70–100	4,482,110	7.9%
101–160	2,481,543	4.4%
161–300	2,316,209	4.1%
301+	894,806	1.6%

Hard-ambiguity coverage

term	rows
de_di_de	22,485,533
shi	15,874,452
zai	7,761,233
yao	5,473,893
ta	3,325,425
zuo	3,081,187
jiao	1,659,343
xian	1,513,147
shijian	797,283
qishi	430,447
changshi	58,936
hangxing	39,425

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

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: 4,191,092 (7.38%).
• Context sentence/turn boundary alignment: 100%.
• Retirement cuts: 1,894,699 (all sentence-final).
• Intra-chunk retirement share: 27.19%.
• Conversational single-turn share: 99.45%.

Mixed-mode & noise

• Mixed tail rows: 16,019,818 · mid-text edit rows: 3,291,800 (edit share 17.05%).
• Noisy rows (any injected edit): 5,697,476 (10.0%).
• ü→v vs ü→u: u-form 566,155 · v-form 5,098,493 (u-share 10.0%).
• All noise-free mid_syllable targets with a Latin tail end in a genuinely incomplete final syllable.

Slicing & multiplicity

• Slice rows: 7,017,910 (cut from long formal entries at sentence-preferred boundaries).
• Up to 4 distinct rows per source pool entry (conversational-first assignment).
• 1,191,396 exact prompt+target duplicates dropped.

Register (target char mass by source)

Conversational sources carry 43.0% of target char mass.

source	rows	target chars	char share
chinesewebtext2_hq	7,507,058	467.1M	34.5%
lccc_large	30,858,374	407.2M	30.1%
dureader_retrieval_corpus	4,311,996	260.0M	19.2%
personal_dialog	12,834,156	159.9M	11.8%
all others	1,298,602	59.5M	4.4%

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	30,858,374	mit	https://huggingface.co/datasets/thu-coai/lccc
personal_dialog	12,834,156	other (research-only)	https://huggingface.co/datasets/silver/personal_dialog
chinesewebtext2_hq	7,507,058	apache-2.0	https://huggingface.co/datasets/Morton-Li/ChineseWebText2.0-HighQuality
dureader_retrieval_corpus	4,311,996	apache-2.0	https://huggingface.co/datasets/zyznull/dureader-retrieval-corpus
clapai_sentiment_zh	578,628	apache-2.0	https://huggingface.co/datasets/clapAI/MultiLingualSentiment
somebreeze_news	286,070	apache-2.0	https://huggingface.co/datasets/somebreeze/Chinese-news-summery
lccc_dialogue	154,574	mit (derived)	https://huggingface.co/datasets/thu-coai/lccc
dureader_retrieval_queries	85,120	apache-2.0	https://huggingface.co/datasets/zyznull/dureader-retrieval-corpus
pd_dialogue	88,778	other (research-only)	https://huggingface.co/datasets/silver/personal_dialog
feilongfl_news	54,260	apache-2.0	https://huggingface.co/datasets/feilongfl/ChineseNewsSummary
cmrc2018	51,172	cc-by-sa-4.0	https://huggingface.co/datasets/hfl/cmrc2018

Files

• train-00000.parquet … train-00029.parquet — train split (30 shards, ~349 MB each)
• dev.parquet — dev split (~218 MB)
• test.parquet — test split (~217 MB)
• metadata.json — build metadata, distributions, serving_contract, 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.