maxbeettww/VeriPromise_ESG_2026_9906_SetFit

VeriPromiseESG Multi-Task SetFit

This repository stores SetFit artifacts generated by app/model/model_setfit.ipynb.

Artifact Layout

fold_1/encoder/
fold_1/task_heads.joblib
...
fold_5/encoder/
fold_5/task_heads.joblib
setfit_thresholds.json
setfit_inference_config.json
setfit_oof_predictions.csv
setfit_oof_probabilities.csv

Model

• Primary encoder: yiyanghkust/finbert-tone-chinese
• Fallback encoder: yiyanghkust/finbert-tone-chinese
• One shared SentenceTransformer encoder is trained once per fold.
• Encoder checkpoints maximize competition-weighted T1-T4 validation pair AP.
• Task heads are selected by fold validation macro F1.
• T2 supports an ordinal cumulative-logistic head.
• Inference routing matches the original submission format.

Thresholds

• T1 threshold: 0.21
• T3 threshold: 0.48

Validation

• OOF average macro F1: not recorded
• OOF competition score: not recorded
• T1/T3 quality gate passed: True

Inference

Use app/model/model_setfit_inference.ipynb and pass this repository id to ensemble_inference_and_export.

Repository id: maxbeettww/VeriPromise_ESG_2026_9906_SetFit

Output Columns

id,promise_status,verification_timeline,evidence_status,evidence_quality