SetSUMBT-dst-nlu-multiwoz21

This model is a fine-tuned version SetSUMBT of roberta-base on MultiWOZ2.1. This model is a combined DST and NLU model and is a distribution distilled version of a ensemble of 5 models. This model should be used to produce uncertainty estimates for the dialogue belief state.

Refer to ConvLab-3 for model description and usage.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.00001
• train_batch_size: 3
• eval_batch_size: 16
• seed: 0
• gradient_accumulation_steps: 1
• optimizer: AdamW
• loss: Ensemble Distribution Distillation Loss
• lr_scheduler_type: linear
• num_epochs: 50.0

Framework versions

• Transformers 4.17.0
• Pytorch 1.8.0+cu110
• Datasets 2.3.2
• Tokenizers 0.12.1