cadazar/han2han-rl

Han2Han RL

Han2Han (169M parameters, encoder-decoder) trained with reinforcement learning to convert Korean text between its two scripts, in both directions:

• Hangul to Hanja: restore the Sino-Korean characters of a mixed-script text from its Hangul-only reading (조선미술 to 朝鮮美術).
• Hanja to Hangul: write a mixed-script text in Hangul only.

It is the last of three checkpoints: cadazar/han2han-pt (pre-trained), cadazar/han2han-it (instruction-tuned), and this one, which starts from han2han-it and takes its prompts the same way. The paper (Findings of EMNLP 2026) reports on the first two; this checkpoint is follow-up work, and its sentence embeddings have not been evaluated.

What to expect

On 256 sentences from 1920s newspapers, held out from the reinforcement learning, greedy decoding:

	han2han-it	this model
Hangul to Hanja: share of the original Hanja restored	0.31	0.63
sentences restored without any error	1%	6%
Hanja to Hangul: share of the Hanja read correctly	0.89	0.93
sentences transcribed without any error	34%	50%
outputs that end on their own (to Hanja / to Hangul)	99% / 100%	100% / 100%
three held-out articles in slices of 250 to 500 characters: Hanja restored	0.81	0.86

Restoring Hanja is the hard direction: the Hangul reading does not say which of several same-sounding characters was meant, and the model has to choose from context. Most restored sentences therefore contain at least one wrong character, and the output needs checking against a source before it is quoted. The errors that remain:

• Homophones: a character with the right sound and the wrong meaning (轉死 for 轉寫, 情書 for 情緖).
• Proper names, which context does not decide (綠香展 for 綠鄕展).
• Words left in Hangul (단순 for 單純).
• Substituted words: occasionally a different word in place of the original (創作 for 習作), which changes the reading as well.

Sampling at temperature 0.6 scores below greedy decoding in both directions (0.60 and 0.91). A second training run with another seed and another held-out split reached 0.67 and 0.92, so differences of a few points between runs are noise.

The newspapers these sentences come from were part of pre-training, so the text is held out from the fine-tuning stages, not unseen by the model.

Usage

Requirements: torch and transformers (tested with torch 2.14.1 and transformers 5.18.0 on CPU). trust_remote_code=True is needed because the model class ships in this repo.

The task prompt goes in the system message and the text alone in the user message, as in han2han-it. The reinforcement learning used one wording per direction; use them as written.

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "cadazar/han2han-rl"
tokenizer = AutoTokenizer.from_pretrained(repo, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(repo, trust_remote_code=True).eval()

TO_HANJA = "한글을 한자로 전사하시오:"
TO_HANGUL = "한자를 한글로 전사하시오:"


def convert(task_prompt, text):
    messages = [
        {"role": "system", "content": task_prompt},
        {"role": "user", "content": text},
    ]
    inputs = tokenizer.apply_chat_template(
        messages, add_generation_prompt=True, return_tensors="pt", return_dict=True
    )
    budget = 2 * len(tokenizer(text).input_ids) + 8
    with torch.no_grad():
        output = model.generate(**inputs, max_new_tokens=budget, do_sample=False)
    return tokenizer.decode(output[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True)


hangul = (
    "단순한 스케치나 카메라적 전사가 아니라 작가의 예술적 정서의 전체적인 안계로 "
    "이루어지는 가장 사실적인 표현을 말하는 것이다."
)
hanja = convert(TO_HANJA, hangul)
print(hanja)
# 단순한 스케치나 카메라的 轉死가 아니라 作家의 藝術的 情書의 全體的인 眼界로
# 이루어지는 가장 事實的인 表現을 말하는 것이다.
print(convert(TO_HANGUL, hanja))
# 단순한 스케치나 카메라적 전사가 아니라 작가의 예술적 정서의 전체적인 안계로
# 이루어지는 가장 사실적인 표현을 말하는 것이다.

The outputs are shown as generated (each is one line). The original sentence reads 單純한 스케치나 카메라的 轉寫가 아니라 作家의 藝術的 情緖의 全體的인 眼界로 이루어지는 가장 寫實的인 表現을 말하는 것이다.: the model left 단순 in Hangul and chose same-sounding characters for three words (轉死, 情書, 事實的). han2han-it returns this sentence in Hangul, unchanged.

• Length. The reinforcement learning used single sentences of 10 to 120 characters. Longer input works (the first 478 characters of a 1922 article come back with 84% of their Hanja, the same as han2han-it), but for long texts it is safer to split at sentence boundaries into pieces of a few hundred characters and convert each piece.
• Token budget. A correct answer needs at most about 1.7 times the tokens of its input, so 2 * input tokens + 8 leaves room and stops a runaway output.
• Thinking. The chat template accepts enable_thinking=True, but neither this model nor han2han-it learned a separate reasoning step; leave it off.
• Other tasks. Rows of the instruction-tuning data were replayed during the reinforcement learning to keep its other behavior in place (the loss on held-out rows of that data went from 2.39 to 2.11), but only transcription was evaluated.

Serving

The model works with the transformers server and its chat client:

transformers serve --trust-remote-code
transformers chat cadazar/han2han-rl --system-prompt "한글을 한자로 전사하시오:"

AutoModelForCausalLM loads Han2HanForCausalLM, a wrapper that takes the rendered chat prompt as one sequence and returns the prompt followed by the reply. AutoModelForSeq2SeqLM loads the encoder-decoder itself; the han2han-it model card describes both, the prompt format, and the tokenizer.

Training

GRPO from han2han-it (main), 300 steps, all weights trained in fp32:

• Prompts: 39,495 newspaper sentences from the training split of the sentence-level transcription set, half of the prompts in each direction, 16 prompts with 8 answers each per step, sampled at temperature 0.7, learning rate 5e-6. The pool was deduplicated against the validation and test splits, which hold the evaluation sentences above and the paper's transcription evaluation, so the model was not trained on either.
• Rewards are rules, with no judge model: the output ends on its own, has the length of its input, restores the original Hanja (or reads them correctly), reads back to its input, and does not repeat itself.
• Keeping the rest in place: a KL penalty to the starting model, and a supervised loss on replayed instruction-tuning rows at half weight.

The reinforcement-learning code is not in the GitHub repo yet.

Files

File	Contents
model.safetensors	fp32 weights, 169.2M parameters, plus the jbu / cbu subword bucket tables
config.json, generation_config.json	model and generation config, with auto_map entries for the Auto classes
tokenizer.json, tokenizer_config.json, chat_template.jinja	fast tokenizer (38400 pieces) and chat template, the same as in han2han-it
spiece.model	the SentencePiece model tokenizer.json was built from
modeling_han2han.py, han2han_config.py	modeling code from the GitHub repo at commit af1330e, the same files as in han2han-it

Citation

@inproceedings{han2han2026,
  title     = {Han2Han: Efficient Language-Specific Character Representation
               through Script-Aware Pre-Training for Historical Text Analysis},
  author    = {Adams, Cellik and Jo, EunKyoung and Kim, Juae},
  booktitle = {Findings of the Association for Computational Linguistics: EMNLP 2026},
  year      = {2026}
}

License

Apache License 2.0, the same as the GitHub repo.