BGE Subregister Classifier Heads
The release contains 17 separately trained logistic-regression heads. Each consumes 1,024-dimensional input vectors. It does not contain a newly trained BGE text encoder.
Research release
Heads and pipeline components
The loader returns class probabilities without applying a scaler. Some folders also contain scaler.pkl files; those are pipeline components, not separate model releases.BGE subregister classifiers: loading and prediction code
Training metadata also contain empty-string class labels. Their meaning is not explained here as a named text category.en_ID: classifier training detailsen_NA-nb: classifier training details
English heads
en_ID predicts the published classes sports. Its training log is dated 26 September 2025.en_ID: classifier training details
en_NA-nb predicts the published classes comments. Its training log is dated 26 September 2025.en_NA-nb: classifier training details
en_NA-nb-OP predicts the published classes culture, dining and lifestyle. Its training log is dated 26 September 2025.en_NA-nb-OP: classifier training details
Finnish heads
fi_ID predicts the published classes sports. Its training log is dated 26 September 2025.fi_ID: classifier training details
fi_ID-NA predicts the published classes comments. Its training log is dated 26 September 2025.fi_ID-NA: classifier training details
fi_NA-nb predicts the published classes comments. Its training log is dated 26 September 2025.fi_NA-nb: classifier training details
fi_NA-nb-OP predicts the published classes culture and consumption. Its training log is dated 26 September 2025.fi_NA-nb-OP: classifier training details
fi_NA-nb-OP-rv predicts the published classes books, dining, beverages and cosmetics. Its training log is dated 26 September 2025.fi_NA-nb-OP-rv: classifier training details
Swedish heads
sv_HI-NA-nb-re predicts the published classes crafts. Its training log is dated 26 September 2025.sv_HI-NA-nb-re: classifier training details
sv_ID predicts the published classes sports and help. Its training log is dated 26 September 2025.sv_ID: classifier training details
sv_ID-NA-nb predicts the published classes comments. Its training log is dated 26 September 2025.sv_ID-NA-nb: classifier training details
sv_IN-NA-nb predicts the published classes organizations. Its training log is dated 26 September 2025.sv_IN-NA-nb: classifier training details
sv_NA-nb predicts the published classes comments. Its training log is dated 26 September 2025.sv_NA-nb: classifier training details
sv_NA-nb-OP predicts the published classes finance. Its training log is dated 26 September 2025.sv_NA-nb-OP: classifier training details
sv_NA-nb-OP-rv predicts the published classes lifestyle and culture. Its training log is dated 26 September 2025.sv_NA-nb-OP-rv: classifier training details
sv_NA-ob-OP predicts the published classes sports and general. Its training log is dated 26 September 2025.sv_NA-ob-OP: classifier training details
sv_ds-IP-NA-nb predicts the published classes travel and contests. Its training log is dated 26 September 2025.sv_ds-IP-NA-nb: classifier training details
Preserve the embedding input
Practical guidance. Establish the exact text encoder before deploying the heads. Matching vector dimensions is not enough to make different embedding spaces compatible. Keep the selected head and its class-name metadata together when interpreting probabilities.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Research release |
| Access | Downloadable |
| Model identifiers | bge-embeddings-subregister-classification |
| License | Apache 2.0 in repository metadata.BGE subregister classifiers: trained artifacts and license metadata |
| Input | Precomputed vectors with 1,024 dimensions.BGE subregister classifiers: loading and prediction code |
| Artifact type | Trained model.pkl files plus metadata and loading code.BGE subregister classifiers: loading and prediction code |
| Encoder identity | The reviewed loader does not identify an exact upstream BGE encoder revision.BGE subregister classifiers: loading and prediction code |
Related models
- BGE Multilingual Web Register Classifier
TurkuNLP ยท Web register classification
This register classifier adapts BGE-M3 RetroMAE and publishes a specific trained cross-validation fold.
Launch date not confirmedDownloadable