distilroberta-base fine-tuned on GoEmotions (27 emotions + neutral, multi-label) with knowledge distillation
from two roberta-large teachers (bce,
asl).
exported to ONNX for transformers.js. onnx/model_quantized.onnx is int8 dynamic quantization
(79 mb), onnx/model_fp16.onnx is for webgpu, onnx/model.onnx is fp32.
test set (int8 model, per-label thresholds tuned on validation): macro-f1 0.543, micro-f1 0.612, ekman macro-f1 0.646. cpu latency about 9 ms per sentence.
the outputs are logits, use sigmoid(logits / 0.720) for calibrated probabilities. the tuned
thresholds are in calib.json. live demo: https://nlp-cv-lab.vercel.app/p/emotion-classification,
code: https://github.com/manasrpandya/NLP_CV_Lab/tree/main/projects/06-emotion-classification
onnx/model.onnx 329 MBonnx/model_fp16.onnx 164 MBonnx/model_quantized.onnx 83 MB