README.md

metadata

language:
  - en
license: apache-2.0
tags:
  - t5-small
  - text2text-generation
  - natural language generation
  - conversational system
  - task-oriented dialog
datasets:
  - ConvLab/tm1
  - ConvLab/tm2
  - ConvLab/tm3
metrics:
  - Slot Error Rate
  - sacrebleu
model-index:
  - name: t5-small-nlg-tm1_tm2_tm3
    results:
      - task:
          type: text2text-generation
          name: natural language generation
        dataset:
          type: ConvLab/tm1, ConvLab/tm2, ConvLab/tm3
          name: tm1+tm2+tm3
          split: test
        metrics:
          - type: Slot Error Rate
            value: 2.1
            name: SER
          - type: sacrebleu
            value: 51.5
            name: BLEU
widget:
  - text: |-
      tm1: [inform][pizza_ordering]([name.store][Domino's])

      system: 
  - text: >-
      tm2: [inform][restaurant-search]([name.restaurant][Via 313, the Violet
      Crown Social Club],[price_range][$8 per slice])


      system: 
  - text: |-
      tm3: [inform][movie]([name.movie][Star Wars],[name.movie][The Grudge])

      system: 
inference:
  parameters:
    max_length: 100

t5-small-nlg-sgd

This model is a fine-tuned version of t5-small on 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: 64
• eval_batch_size: 64
• seed: 42
• gradient_accumulation_steps: 8
• total_train_batch_size: 512
• 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