Qing official-title decomposition adapter
A LoRA adapter for Qwen/Qwen3.5-4B. It separates compound Qing bureaucratic titles into fifteen institutional components. The accompanying code converts components to ranks separately; the model does not predict numeric ranks or career mobility.
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: https://github.com/bruceyyu/decompose-chinese-official-titles
Use
Install the codebase with pip install -e '.[inference]', then:
official-titles parse --task qing --adapter /path/to/adapter \
--input my-titles.jsonl --output my-components.jsonl --with-ranks
Each input is a JSON string or an object with original_title. The exact
instruction is in prompt.txt. Disable thinking, decode greedily, and allow
up to 512 new tokens. duty_office is the office field throughout the schema,
instruction, and interface. Single-valued components are strings or null;
multi-valued components are arrays, empty when absent. Output must retain
original_title unchanged.
Adaptation and selection
The adaptation pool contains 1,967 distinct annotated titles: 1,770 for optimization and 197 for internal validation. Two external holdouts each contain 200 titles and are excluded from optimization and checkpoint selection.
The recipe uses rank-64 LoRA, learning rate 2e-4 with linear decay over a six-epoch horizon, four completed epochs, batch size 32, sequence limit 3,000, and shuffle seed 42. Loss applies only to the assistant response and is normalized per example. Selection minimizes internal validation negative log-likelihood at completed epochs. Training configuration is provided; training scripts and reference annotations are not distributed.
Internal validation selected epoch 2, step 110 (NLL 0.025720). This is a new checkpoint from the retraining run, not the unavailable historical checkpoint with the same step number. The selected weights are backed up locally.
Evaluation
On the two 200-title holdouts, the retrained release model scores 94.02% (fixed) and 93.38% (rare-component-stratified), averaging exact agreement for strings and set F1 for arrays across fifteen components. Whole-record exact agreement is 48.0% and 49.0%, respectively. Both sets have 100% schema validity. The input title is reproduced exactly in 99.5% and 100% of cases; the local parser rejects a response that changes the input title.
Scope and limitations
The model is only intended for research on historical bureaucratic titles, especially Jinshenlu. The parser can normalize institutional wording rather than only copy spans. Valid JSON or a recognized rank does not establish a correct institutional reading. Unfamiliar sources, OCR errors, different periods, and new organizations require validation. Additional titles, eligibility, and actual posts are not interchangeable.
The compatible PEFT export preserves separate adapter contributions to the output head. Do not merge it into tied base weights without validating the merged runtime. Base-model weights are obtained separately.
Files and licensing
The model folder contains adapter weights and configuration, this card, prompt,
schema, checksums, and verification information. It contains no base weights,
credentials, human annotation rows, or person-level records. The authors'
adapter contributions and model materials use CC BY-NC 4.0, requiring
attribution and noncommercial use subject to LICENSE. Retain NOTICE.
Qwen's base model retains Apache-2.0 (BASE_MODEL_LICENSE); the adapter terms
do not relicense it. Code 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}
}
- Downloads last month
- -