Translation

wmt20_en_ja

This dataset is created as a test set for the WMT20 shared task on news translation. This is English to Japanese translation.

References:

• Data Source
• 2020 Fifth Conference on Machine Translation (WMT20)

  local dataset = {
    class_path: 'SacreBleuDataset',
    init_args: { name: 'wmt20', langpair: 'en-ja' },
  };
  
  {
    class_path: 'Generation',
    init_args: {
      eval_dataset: dataset,
      few_shot_generator: {
        class_path: 'RandomFewShotGenerator',
        init_args: {
          // Use the eval dataset for few-shot data,
          // but `RandomFewShotGenerator` will avoid using the same few-shot isntances as the input.
          dataset: dataset,
          num_shots: 4,
        },
      },
      prompt_template: |||
        {% for item in few_shot_data %}
        En: `{{ item.source }}`
        Ja: `{{ item.references[0] }}`
        {% endfor %}
        En: `{{ source }}`
      ||| + 'Ja: `',
      metrics: [
        { class_path: 'BLEU', init_args: { tokenize_option: 'ja-mecab' } },
      ],
      gen_kwargs: { max_new_tokens: 128, stop_sequences: ['`'] },
      batch_size: 4,
    },
  }
  

wmt20_ja_en

This dataset is created as a test set for the WMT20 shared task on news translation. This is Japanese to English translation.

References:

• Data Source
• 2020 Fifth Conference on Machine Translation (WMT20)

  local dataset = {
    class_path: 'SacreBleuDataset',
    init_args: { name: 'wmt20', langpair: 'ja-en' },
  };
  
  {
    class_path: 'Generation',
    init_args: {
      eval_dataset: dataset,
      few_shot_generator: {
        class_path: 'RandomFewShotGenerator',
        init_args: {
          // Use the eval dataset for few-shot data,
          // but `RandomFewShotGenerator` will avoid using the same few-shot isntances as the input.
          dataset: dataset,
          num_shots: 4,
        },
      },
      prompt_template: |||
        {% for item in few_shot_data %}
        Ja: `{{ item.source }}`
        En: `{{ item.references[0] }}`
        {% endfor %}
        Ja: `{{ source }}`
      ||| + 'En: `',
      metrics: [
        { class_path: 'BLEU', init_args: { tokenize_option: 'intl' } },
      ],
      gen_kwargs: { max_new_tokens: 128, stop_sequences: ['`'] },
      batch_size: 4,
    },
  }