Update: V2+ contains new weights for the new classification (2026).
V1: Repository contains models used in the paper: Beręsewicz, M., Wydmuch, M., Cherniaiev, H., & Pater, R. (2026). Multilingual Hierarchical Classification of Job Advertisements for Job Vacancy Statistics. Journal of Official Statistics, 42(1), 23–61. https://doi.org/10.1177/0282423X251395400
How to use: https://github.com/OJALAB/paper-job-ads-classifier
Models used in the paper (2024)
Linear models:
- linear-bottom-2024.zip (136.2 MB)
- linear-top-2024.zip (138.4 MB)
Transformer models based on Polish data (trained with HerBERT-base and HerBERT-large):
- transformer-bottom-base-2024.zip (1.6 GB)
- transformer-bottom-large-2024.tar.gz (split into .partaa and .partab; 4.8 GB in total)
- transformer-top-base-2024.zip (1.6 GB)
- transformer-top-large-2024.tar.gz (split into .partaa and .partab; 4.8 GB in total)
Transformer models based on Polish data (trained with xlm-RoBERTa-base and xlm-RoBERTa-large):
- transformer-bottom-base-rob-2024.zip (2.3 GB)
- transformer-bottom-large-rob-2024.tar.gz (split into .partaa and .partab; 5.7 GB in total)
- transformer-top-base-rob-2024.zip (2.3 GB)
- transformer-top-large-rob-2024.tar.gz (split into .partaa and .partab; 5.7 GB in total)
Transformer models based on multilingual data (Polish and English; trained with xlm-RoBERTa-base only):
- transformer-bottom-rob-2024-argos.zip (2.4 GB): dataset translated with Argos
- transformer-bottom-rob-2024-google.zip (2.4 GB): dataset translated with Google Translate
- transformer-top-rob-2024-argos.zip (2.4 GB): dataset translated with Argos
- transformer-top-rob-2024-google.zip (2.4 GB): dataset translated with Google Translate
Transformer models based on multilingual data (24 languages; trained with xlm-RoBERTa-base only):
- transformer-multi-bottom-2024.zip (2.8 GB)
- transformer-multi-top-2024.zip (2.0 GB)
Excel file with information about the quality of classifications:
- models-accuracy.xlsx (368.2 KB)
Models for the new classification (2026)
Transformer models based on Polish data (trained with HerBERT-base and ModernBERT-base):
- transformer-bottom-base-her-2026.zip (4.1 GB) [new in V3]
- transformer-top-base-her-2026.zip (4.1 GB) [new in V3]
- transformer-bottom-base-mod-2026.zip (split into .zip_partaa and .zip_partab; 5.0 GB in total)
- transformer-top-base-mod-2026.zip (split into .zip_partaa and .zip_partab; 5.0 GB in total)
Transformer models based on multilingual data (Polish and English, dataset translated with Google Translate; trained with xlm-RoBERTa-base and [MM]):
- transformer-bottom-rob-2026-google.zip (split into .zip_partaa and .zip_partab; 5.4 GB in total)
- transformer-top-rob-2026-google.zip (split into .zip_partaa and .zip_partab; 5.4 GB in total)
- transformer-bottom-mm-2026-google.zip (split into .zip_partaa and .zip_partab; 7.4 GB in total)
- transformer-top-mm-2026-google.zip (split into .zip_partaa and .zip_partab; 7.4 GB in total)
Excel file with information about the quality of classifications (test data only):
- models-accuracy-2026-test-data-only.xlsx (16.0 KB)
Reassembling split archives
Download all parts into one directory, concatenate them, then extract:
cat transformer-top-large-2024.tar.gz.part* > transformer-top-large-2024.tar.gz
tar -xzf transformer-top-large-2024.tar.gz
cat transformer-top-base-mod-2026.zip_part* > transformer-top-base-mod-2026.zip
unzip transformer-top-base-mod-2026.zip
On Windows: copy /b file.zip_partaa + file.zip_partab file.zip
(2026)