README.md

metadata

language:
  - en
license: apache-2.0
tags:
  - t5-small
  - text2text-generation
  - natural language understanding
  - conversational system
  - task-oriented dialog
datasets:
  - ConvLab/tm3
metrics:
  - Dialog acts Accuracy
  - Dialog acts F1
model-index:
  - name: t5-small-nlu-tm3-context3
    results:
      - task:
          type: text2text-generation
          name: natural language understanding
        dataset:
          type: ConvLab/tm3
          name: Taskmaster-3
          split: test
          revision: 910584e5451e2e439bb2a07b8544ecb42ff8835b
        metrics:
          - type: Dialog acts Accuracy
            value: 89
            name: Accuracy
          - type: Dialog acts F1
            value: 85.1
            name: F1
widget:
  - text: >-
      system: OK. And where will you be seeing the movie?

      user: In Creek's End, Oregon

      system: Creek’s End, Oregon. Got it. Is there a particular movie you have
      in mind?

      user: Mulan, please. We are taking the kids
  - text: >-
      system: No problem. It looks like tonight’s remaining showtimes for Mulan
      at AMC Mercado 24 are 5:00pm, 7:10pm, and 9:45pm. Which is best for you?

      user: I would like the earliest time, 5:00pm

      system: Great. And how many tickets?

      user: three please
inference:
  parameters:
    max_length: 100

t5-small-nlu-tm3-context3

This model is a fine-tuned version of t5-small on Taskmaster-3 with context window size == 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