README.md

metadata

language:
  - en
license: apache-2.0
tags:
  - t5-small
  - text2text-generation
  - natural language understanding
  - conversational system
  - task-oriented dialog
datasets:
  - ConvLab/multiwoz21
  - ConvLab/sgd
  - ConvLab/tm1
  - ConvLab/tm2
  - ConvLab/tm3
metrics:
  - Slot Error Rate
  - sacrebleu
widget:
  - text: >-
      multiwoz21: user: I would like a taxi from Saint John's college to Pizza
      Hut Fen Ditton.
    example_title: MultiWOZ 2.1
  - text: >-
      sgd: user: Could you get me a reservation at P.f. Chang's in Corte Madera
      at afternoon 12?
    example_title: Schema-Guided Dialog
  - text: 'tm1: user: I would like to order a pizza from Domino''s.'
    example_title: Taskmaster-1
  - text: 'tm2: user: I would like help getting a flight from LA to Amsterdam.'
    example_title: Taskmaster-2
  - text: >-
      tm3: user: Well, I need a kids friendly movie. I was thinking about seeing
      Mulan.
    example_title: Taskmaster-3
inference:
  parameters:
    max_length: 100
base_model: t5-small
model-index:
  - name: t5-small-nlu-multiwoz21_sgd_tm1_tm2_tm3
    results:
      - task:
          type: text2text-generation
          name: natural language understanding
        dataset:
          name: MultiWOZ 2.1
          type: ConvLab/multiwoz21
          split: test
          revision: 5f55375edbfe0270c20bcf770751ad982c0e6614
        metrics:
          - type: Dialog acts Accuracy
            value: 77.5
            name: Accuracy
          - type: Dialog acts F1
            value: 86.4
            name: F1
      - task:
          type: text2text-generation
          name: natural language understanding
        dataset:
          name: SGD
          type: ConvLab/sgd
          split: test
          revision: 6e8c79b888b21cc658cf9c0ce128d263241cf70f
        metrics:
          - type: Dialog acts Accuracy
            value: 45.2
            name: Accuracy
          - type: Dialog acts F1
            value: 58.6
            name: F1
      - task:
          type: text2text-generation
          name: natural language understanding
        dataset:
          name: TM1+TM2+TM3
          type: ConvLab/tm1, ConvLab/tm2, ConvLab/tm3
          split: test
        metrics:
          - type: Dialog acts Accuracy
            value: 81.8
            name: Accuracy
          - type: Dialog acts F1
            value: 73
            name: F1

t5-small-nlu-multiwoz21_sgd_tm1_tm2_tm3

This model is a fine-tuned version of t5-small on MultiWOZ 2.1, Schema-Guided Dialog, Taskmaster-1, Taskmaster-2, and Taskmaster-3.

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