review -> PhoBERT
       -> nested-OOF RF probabilities
       -> nested-OOF predicted helpfulness_score/confidence
       -> safe metadata
       -> final classifier

Raw helpfulness_score and confidence are annotation outputs and are not used as inference inputs. They only supervise auxiliary predictions and sample weights during training.

Best cross-validation model

phobert_metadata_logreg

See kfold_metrics_summary.csv, oof_predictions.csv, training_report.json and config.json for full results.

DPhO05/tiki-review-helpfulness-phobert-berf-gold-v2

Tiki Review Helpfulness — PhoBERT BERF Gold v2

Dataset

Leakage-safe pipeline

Best cross-validation model

DPhO05/
tiki-review-helpfulness-phobert-berf-gold-v2