cadazar/han2han-rl-data

Han2Han RL data

The open data subset of Han2Han's GRPO stage is a pool of Hanja-Hangul mixed sentences for the transcription tasks and a set of instruction-following chat turns for the replay term. The released run (cadazar/han2han-rl) trained on data that cannot be redistributed, but this repository holds what a public re-run of that stage can read. Everything here is derived from other people's datasets under their licenses, and the whole set is published under CC BY-NC 4.0, the most restrictive of them.

Files

file	rows	content
pool/train.parquet	40,000	training sentences
pool/eval.parquet	1,024	evaluation sentences from documents that contribute no training sentence; the trainer draws its 256 held-out sentences from them by seed
replay/train.parquet	4,000	replay chat turns (open-korean-instructions)
replay/validation.parquet	400	chat-NLL probe turns (open-korean-instructions)

rl/build_public_data.py in the han2han-v1-public repository builds every file, and its komagpie command adds the KoMagpie half of the replay set on your machine (see below).

Pool

The pool's source is the Open Korean Historical Corpus (OKHC, CC BY-NC 4.0), three of its corpora, read from the Hub repository's files news_archive_part_001_of_003.jsonl, news_archive_part_002_of_003.jsonl, news_archive_part_003_of_003.jsonl, magazine.jsonl, newslibrary_part_001_of_150.jsonl, newslibrary_part_002_of_150.jsonl, newslibrary_part_003_of_150.jsonl, newslibrary_part_004_of_150.jsonl, newslibrary_part_005_of_150.jsonl, newslibrary_part_006_of_150.jsonl, newslibrary_part_007_of_150.jsonl, newslibrary_part_008_of_150.jsonl, newslibrary_part_009_of_150.jsonl and newslibrary_part_010_of_150.jsonl.

The columns are target (the mixed-script sentence as the archive prints it), transcription, year (int), source (the archive), corpus, okhc_id (the OKHC document id), url (the document's page at the archive) and copyright (the OKHC field, always "Public Domain").

transcription is the rule-based Hangul reading han2han_tools.transcribe produces from target, the same reading the RL rewards compute, but it is not a gold reading.

Copyright field per corpus (the files read)

corpus	value	documents
Korean Newspaper Archive	Public Domain	364,409
Korean Modern and Contemporary Magazines (한국근현대잡지자료)	Public Domain	15,326
Naver News Library	Public Domain	231,415
Naver News Library	All Rights Reserved	670,922

Only documents marked "Public Domain" are read.

Filters, in order, with counts

stage	count
documents in the files	1,282,072
not Public Domain, dropped	670,922
Public Domain	611,150
no year, dropped	1,091
year before 1890, dropped	3,202
documents from 1890 on	606,857
sentences after splitting	2,832,865
outside 10 to 120 characters, dropped	1,167,531
no Hangul or Hanja ratio under 0.15, dropped	911,452
archive marks (lacunae, unreadable marks, archaic Hangul, PUA, kana), dropped	206,156
clean sentences	547,726
duplicates under the whitespace-stripped key, dropped	12,313
deduplicated	535,413
transcription of another length, dropped	0
sentence fails the tokenizer round trip, dropped	19,682
transcription fails the tokenizer round trip, dropped	0
after the transcription and tokenizer checks	515,731 (in 131,736 documents)
documents the private used-source ledger already holds, dropped	3,699 documents, 29,265 sentences
sentences equal to a CASASIA evaluation sentence, dropped	3
sentences equal to a sentence of a CASASIA article, dropped	0
sentences inside the demo article, dropped	1
after the leakage checks	486,462 (in 128,037 documents)

The sentence rule is the one the private transcription SFT build used. A sentence ends at terminal punctuation (. ! ? 。 ！ ？ …) followed by whitespace, or at a newline. The length, Hangul and Hanja-ratio filters and the dedup are rl/clean_pool.py's (Hanja counted over U+4E00 to U+9FFF), and the transcription-length and tokenizer round-trip checks and the archive-mark screens are rl/data.py's (the han2han_v2_tokenizer of the code repository).

Draw

Documents are shuffled with seed 0. Whole documents go to the eval set until it holds 1,024 sentences, and the remaining documents fill each corpus's quota in shuffled order, a document's sentences together. Quotas are proportional to the usable sentences per corpus with no corpus above 50% of the training pool (the newspaper corpora have far more usable sentences than the magazines).

corpus	usable sentences	train sentences	train documents	years
Naver News Library	301,175	20,000	7,471	1920 to 1962
Korean Modern and Contemporary Magazines (한국근현대잡지자료)	176,624	19,067	836	1896 to 1942
Korean Newspaper Archive	8,663	933	925	1898 to 1952
all	486,462	40,000	9,232	1896 to 1962

The eval set holds 848 sentences in 327 documents of the Naver News Library, 21 sentences in 21 documents of the Korean Newspaper Archive and 155 sentences in 12 documents of the Korean Modern and Contemporary Magazines (한국근현대잡지자료). Its years run from 1903 to 1962. The public recipe runs the trainer with --data_path pool/eval.parquet --n_eval 256 --train_pool pool/train.parquet. The trainer then draws 256 sentences by seed from this file and the training prompts from the pool, mirroring the private protocol.

Leakage checks

The checks were run once against the paper's evaluation data on the private side, and nothing of it is shipped. The used-source ledger (every document Han2Han's synthetic SFT, RL pools and held-out evaluations already used, matched by text shingles and URL) removed 3,699 documents holding 29,265 sentences. The CASASIA sentence file removed 3 sentences, the sentences of the CASASIA articles 0, and the demo article 1.

Replay

The columns are messages (a list of {role, content} with a system, a user and an assistant turn), bucket (always instruction_following) and source.

Each row is one single-turn instruction-following example framed as Han2Han's IT stage framed it. The system turn is the row's own instruction when it has one, otherwise one of the five generic prompts the training collator samples ("당신은 도움이 되는 한국어 어시스턴트입니다.", "주어진 질문에 정확하고 친절하게 답하시오.", "다음 작업을 수행하시오.", "사용자의 요청에 신중히 응답하시오.", "다음 질문에 답하시오."), the user turn is the input text and the assistant turn the target. Rows with an empty input or target are skipped.

source	upstream	license	train rows	validation rows
open_kor_instructions	heegyu/open-korean-instructions	MIT (its five constituents carry their own terms, below)	4,000	400
komagpie	channelcorp/KoMagpie-raw	none stated on the Hub; not shipped	0 (4,000 when fetched)	0 (400 when fetched)

The open-korean-instructions rows are drawn with seed 0 from the single-turn rows of the upstream (one <usr> and one <bot> tag, at most one <sys>). The draw holds 268 training rows with their own instruction and 3,732 with a generic prompt (30 and 370 in validation), and 0 rows with an empty input or target were skipped. The upstream is a merge of KoAlpaca v1.0 (Bingsu/ko_alpaca_data, CC BY-NC 4.0), KoAlpaca v1.1 (beomi, no license tag on the Hub), ShareGPT DeepL ko (junelee/sharegpt_deepl_ko, no license tag on the Hub), OIG-small-chip2-ko (heegyu, Apache 2.0) and korquad-chat-v1 (heegyu, MIT).

KoMagpie rows are not in this repository. The upstream, channelcorp/KoMagpie-raw, states no dataset license (its card only links the license of the model that wrote its answers), so it is not redistributed here. The released run's replay mix held 4,000 training and 400 validation rows of it. To add them, run python rl/build_public_data.py komagpie --out <this directory> from the code repository. The command downloads the upstream shards, prepares them as the private build did (empty rows dropped, shuffled with seed 42, 0.2% validation), draws 4,000 and 400 rows with seed 0 and writes replay/komagpie_train.parquet and replay/komagpie_validation.parquet in the layout above, to concatenate with the shipped files.

Differences from the released run

The released cadazar/han2han-rl trained on a pool cut from the paper's transcription SFT source and was evaluated on 256 held-out CASASIA sentences, neither redistributable, and its replay mix also held AI Hub summaries and article transcriptions. This repository's pool comes from the OKHC instead, its evaluation sentences come from the same corpora (documents disjoint from the pool), and its replay set holds the open-korean-instructions turns alone (the builder's komagpie command adds KoMagpie-raw on your machine).

Reference numbers

The numbers below come from greedy decoding with the public repo's rl/jax/evaluate.py (it layout, 256 sentences of pool/eval.parquet drawn by seed 0, each in both directions, budget 2x + 8 tokens) on the two released Flax checkpoints.

model	direction	terminates	hanja_recall	reading_recall	round_trip	exact
cadazar/han2han-it-flax (762)	Hangul to Hanja	0.996	0.654	-	0.970	0.113
cadazar/han2han-it-flax (762)	Hanja to Hangul	1.000	-	0.952	0.973	0.422
cadazar/han2han-rl-flax (300)	Hangul to Hanja	1.000	0.831	-	0.977	0.145
cadazar/han2han-rl-flax (300)	Hanja to Hangul	1.000	-	0.977	0.987	0.609

The demo article (rl/demo_article.txt of the public repo, held out of every split) scores a joined recall of 0.810 (IT) and 0.855 (RL) in 120-character chunks, and taken whole it scores 0.845 for both. These sentences are easier than the private held-out set the model cards report (the OKHC rows are cleaner print with fewer glosses), so the card numbers and these are not comparable to each other, but the IT-to-RL gain carries over.

Citations

The pool is derived from the Open Korean Historical Corpus (CC BY-NC 4.0):

@article{song2025open,
  author  = {Seyoung Song and Nawon Kim and Songeun Chae and Kiwoong Park and Jiho Jin and Haneul Yoo and Kyunghyun Cho and Alice Oh},
  journal = {ArXiv preprint},
  title   = {Open Korean Historical Corpus: A Millennia-Scale Diachronic Collection of Public Domain Texts},
  url     = {https://arxiv.org/abs/2510.24541},
  volume  = {abs/2510.24541},
  year    = {2025}
}

The replay rows are derived from heegyu/open-korean-instructions (MIT license, Heegyu Kim, 2023, https://huggingface.co/datasets/heegyu/open-korean-instructions, code at https://github.com/HeegyuKim/open-korean-instructions) and, when fetched, from channelcorp/KoMagpie-raw (Channel Corp., 2024, https://huggingface.co/datasets/channelcorp/KoMagpie-raw, with no dataset license stated), a Korean adaptation of the Magpie method:

@misc{kim2023openkoreaninstructions,
  author = {Heegyu Kim},
  title  = {open-korean-instructions},
  year   = {2023},
  url    = {https://huggingface.co/datasets/heegyu/open-korean-instructions}
}

@misc{channelcorp2024komagpie,
  author = {Channel Corp.},
  title  = {KoMagpie-raw},
  year   = {2024},
  url    = {https://huggingface.co/datasets/channelcorp/KoMagpie-raw}
}

@article{xu2024magpie,
  title   = {Magpie: Alignment Data Synthesis from Scratch by Prompting Aligned LLMs with Nothing},
  author  = {Zhangchen Xu and Fengqing Jiang and Luyao Niu and Yuntian Deng and Radha Poovendran and Yejin Choi and Bill Yuchen Lin},
  journal = {arXiv preprint arXiv:2406.08464},
  year    = {2024}
}