JobBERT-zh initialization

Domain-adapted Chinese recruitment encoder with an inherited CRF checkpoint.

Project · Data · Reproduction guide

Use this model

This package combines a Transformers encoder with a custom CRF. It is not a standard AutoModelForTokenClassification export. Loading the encoder alone does not reproduce span extraction. See loading and technical details for the custom model class and checkpoint paths.

The model predicts flat character spans in Chinese job advertisements using language (L), knowledge (K), occupational-skill (S), and transversal-competence (T) labels. It does not assign ESCO concept IDs or measure individual applicants' abilities.

Checkpoint role

This repository supplies the domain-adapted encoder and CRF used to initialize the later supervision study. The historical exact F1 of 0.4331 uses a hybrid 2,601-record reference and span alignment. The B2 continuation is evaluated under a different human-reference protocol.

File	Purpose
model.safetensors and configuration/tokenizer files	Domain-adapted encoder
crf/best.pt	Inherited CRF checkpoint

Access and scope

The model licence remains other; source-text permissions differ from the software licence. Use for research on recruitment text, with appropriate data permissions and validation. Applicant profiling and automated hiring decisions are outside the intended use.

Recruitment texts in Chinese-SkillSpan were obtained through three routes: purchased collections from MacroData (马克数据网), operated by 重庆马禾锐信息科技有限公司; research-team collection of public recruitment pages across regions of China; and the 招聘数据集 / Recruitment Dataset, ID 163746 on Alibaba Cloud Tianchi. For the public-institution subset, team members manually identified pages, registered selected URLs in a crawling framework that revisited those sources on weekly or monthly schedules, and then screened and cleaned the records. This card does not establish the source composition of the 3M pretraining corpus. See data sources and collection. Source attribution does not add redistribution rights or a promise to supply additional data by email.

Loading instructions, checkpoint information, and evaluation details are in TECHNICAL_DETAILS.md. The resource naming guide maps manuscript names to exact model IDs.