bruceyyu/joseon-title-decomposition

Joseon Korean-script official-title decomposition adapter

A separate LoRA adapter for Qwen/Qwen3.5-4B, trained to identify departments and posts in the Korean-script form of Joseon official titles.

Related manuscript: Office, Rank, and Mobility: Decomposing Historical Chinese Bureaucratic Titles with Language Models, Yue Yu, Jun Chen, Michael Yan Hon Chung, and Cameron Campbell.

Preprint: https://osf.io/preprints/socarxiv/eg87f_v1

Code and installation: https://github.com/bruceyyu/decompose-chinese-official-titles

Use

Install the codebase with pip install -e '.[inference]', download this adapter folder, and run:

official-titles parse --task joseon --adapter /path/to/adapter \
  --input my-korean-titles.jsonl --output my-components.jsonl

Use the exact prompt.txt, disable thinking, decode greedily, and allow up to 256 new tokens. Output has original_title, department, and position. The latter two fields are strings or null.

Source, adaptation, and selection

The source is the third-party JoseonTSI career-record collection: https://github.com/dhchoi02/JoseonTSI . Source files are obtained separately and remain subject to their original terms. Preparation strips parenthetical forms, screens Korean-script titles, deduplicates that form, and constructs 1,000 adaptation and 250 disjoint evaluation titles, using seed 20260930.

Five-fold cross-validation within the 1,000-title adaptation pool selected learning rate 5e-4, four epochs, rank-32 LoRA, batch size 32, shuffle seed 42, maximum sequence length 3,000, and thinking disabled. The released checkpoint is trained on all 1,000 adaptation titles. The 250 evaluation titles are not training or cross-validation inputs.

Evaluation

On the 250-title holdout: department exact agreement 97.2%, position exact agreement 98.0%, component average 97.6%, both components correct 95.2%, and schema validity 100%. These are agreements with the benchmark's source-derived targets, not independent expert verification of all Joseon institutional roles. They do not establish performance on every Korean historical source.

Files and licensing

This package contains only adapter weights/configuration, this card, the prompt, schema, and checksum manifest. It contains no base weights, private labels, credentials, or person-level career records. The authors' adapter contributions and model materials use CC BY-NC 4.0, requiring attribution and noncommercial use subject to the full LICENSE. Retain the attribution in NOTICE. The independently obtained Qwen base model retains Apache-2.0 (BASE_MODEL_LICENSE); the adapter terms do not relicense it. The software uses PolyForm Noncommercial 1.0.0.

Citation

@misc{yu_chen_chung_campbell_2026,
  title={Office, Rank, and Mobility: Decomposing Historical Chinese Bureaucratic Titles with Language Models},
  url={https://osf.io/preprints/socarxiv/eg87f_v1},
  publisher={SocArXiv},
  author={Yu, Yue and Chen, Jun and Chung, Michael Y H and Campbell, Cameron},
  year={2026},
  month={Oct}
}