zhuqi's picture
Update README.md
4498879
|
raw
history blame
2.32 kB
metadata
language:
  - en
license: apache-2.0
tags:
  - t5-small
  - text2text-generation
  - dialogue generation
  - conversational system
  - task-oriented dialog
datasets:
  - ConvLab/multiwoz21
metrics:
  - LM loss
model-index:
  - name: t5-small-goal2dialogue-multiwoz21
    results:
      - task:
          type: text2text-generation
          name: dialogue generation
        dataset:
          type: ConvLab/multiwoz21
          name: MultiWOZ 2.1
          split: validation
          revision: 5f55375edbfe0270c20bcf770751ad982c0e6614
        metrics:
          - type: Language model loss
            value: 1.5253684520721436
            name: LM loss
      - task:
          type: text2text-generation
          name: dialogue generation
        dataset:
          type: ConvLab/multiwoz21
          name: MultiWOZ 2.1
          split: test
          revision: 5f55375edbfe0270c20bcf770751ad982c0e6614
        metrics:
          - type: Language model loss
            value: 1.515929937362671
            name: LM loss
widget:
  - text: >-
      You are traveling to Cambridge and looking forward to try local
      restaurants. You are looking for a particular attraction. Its name is
      called nusha. Make sure you get postcode and address. You are also looking
      for a place to dine. The restaurant should be in the expensive price range
      and should serve indian food. The restaurant should be in the centre. Make
      sure you get address
  - text: >-
      You want to book a taxi. The taxi should go to pizza hut fen ditton and
      should depart from saint john's college. The taxi should leave after
      17:15. Make sure you get car type and contact number
inference:
  parameters:
    max_length: 1024

t5-small-goal2dialogue-multiwoz21

This model is a fine-tuned version of t5-small on MultiWOZ 2.1.

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: 32
  • eval_batch_size: 64
  • seed: 42
  • distributed_type: multi-GPU
  • gradient_accumulation_steps: 4
  • total_train_batch_size: 128
  • 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