bert-base-uncased finetuned on the emotion dataset using PyTorch Lightning. Sequence length 128, learning rate 2e-5, batch size 32, 2 GPUs, 4 epochs.
For more details, please see, the emotion dataset on nlp viewer.
Data came from HuggingFace's
datasets package. The data can be viewed on nlp viewer.
val_acc - 0.931 (useless, as this should be precision/recall/f1)
The score was calculated using PyTorch Lightning metrics.