goemotions-student
distilroberta-base distilled from the two roberta-large teachers (alpha 0.5, t 2) plus the gold labels. pytorch weights of goemotions-distil-onnx.
multi-label: use sigmoid(logits / 0.720) and the per-label thresholds tuned on validation (in the repo's results/metrics.json). test macro-f1 0.534 (95% ci 0.512 to 0.552), micro-f1 0.611, ekman macro-f1 0.646.
code and report: https://github.com/manasrpandya/NLP_CV_Lab/tree/main/projects/06-emotion-classification