FinerWeb Quality Classifier
FinerWeb adapts DeBERTa-v3-base to classify English text lines. Training used 328,472 labelled lines from 20,000 FineWeb documents.
The trained task
Most training labels came from LLM annotation, with a smaller manual agreement sample. The model has nine outputs, but its configuration names them only LABEL_0 to LABEL_8. A published Platt scaler calibrates the clean-class score. The scaler belongs to this pipeline; it is not another language model.FinerWeb Quality Classifier: model card and trained artifactsFinerWeb Quality Classifier: output labels and configuration
Practical assessment
Practical guidance. Use the project pipeline to interpret class indices and apply calibration. A whole page supplied as one line changes the input unit. Check retained and removed lines before using the score to filter a corpus.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | finerweb-quality-classifier |
| Base model | DeBERTa-v3-base.FinerWeb Quality Classifier: model card and trained artifacts |
| Prediction unit | A text line, rather than a whole document.FinerWeb Quality Classifier: model card and trained artifacts |
| License | Apache 2.0 for the model.FinerWeb Quality Classifier: model card and trained artifacts |
Related models
- HPLT3 Persian Quality Classifier
TurkuNLP ยท Web quality classification
This downloadable Persian-named quality classifier has sparse documentation and an XLM-RoBERTa Large ancestry tag.
Launch date not confirmedResearch releaseDownloadable