README.md

metadata

language:
  - en
license: apache-2.0
tags:
  - t5-small
  - text2text-generation
  - natural language understanding
  - conversational system
  - task-oriented dialog
datasets:
  - ConvLab/sgd
metrics:
  - Dialog acts Accuracy
  - Dialog acts F1
model-index:
  - name: t5-small-nlu-sgd
    results:
      - task:
          type: text2text-generation
          name: natural language understanding
        dataset:
          type: ConvLab/sgd
          name: SGD
          split: test
          revision: 6e8c79b888b21cc658cf9c0ce128d263241cf70f
        metrics:
          - type: Dialog acts Accuracy
            value: 45
            name: Accuracy
          - type: Dialog acts F1
            value: 58.6
            name: F1
widget:
  - text: >-
      user: Could you get me a reservation at P.f. Chang's in Corte Madera at
      afternoon 12?
  - text: >-
      user: Sure, may I know if they have vegetarian options and how expensive
      is their food?
inference:
  parameters:
    max_length: 100

t5-small-nlu-sgd

This model is a fine-tuned version of t5-small on Schema-Guided Dialog.

Refer to ConvLab-3 for model description and usage.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.001
• train_batch_size: 128
• eval_batch_size: 64
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 256
• optimizer: Adafactor
• lr_scheduler_type: linear
• num_epochs: 10.0

Framework versions

• Transformers 4.18.0
• Pytorch 1.10.2+cu102
• Datasets 1.18.3
• Tokenizers 0.11.0