metadata

configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
dataset_info:
  features:
    - name: record
      dtype: string
    - name: clean_para_index_set_pair
      dtype: string
    - name: src
      dtype: string
    - name: dst
      dtype: string
    - name: src_text
      dtype: string
    - name: dst_text
      dtype: string
    - name: src_rate
      dtype: float64
    - name: dst_rate
      dtype: float64
  splits:
    - name: train
      num_bytes: 8884444751
      num_examples: 15331650
  download_size: 2443622169
  dataset_size: 8884444751

联合国数字图书馆的段落级中-英对齐平行语料

用我口胡的方法弄出来的平行语料，统计数据和拿argostranslate直接又跑了一份bleu score的结果已经丢论文里了，论文在写了在写了。应该拿这份去练机翻模型没问题，数据源是人写的。

bleu score 这里贴一份吧，懒得转格式了，我不太懂看，可能很差（

Language     & Paragraph Count & Avg Tokens & bleu1 & bleu2 & bleu3 & bleu4 \\
\midrule
ar & 59754 & 52.71873 & 0.73799 & 0.58027 & 0.48118 & 0.40782 \\
de & 187 & 69.58824 & 0.62058 & 0.38837 & 0.26155 & 0.18271 \\
es & 66537 & 50.70776 & 0.74566 & 0.58545 & 0.48445 & 0.41073 \\
fr & 68765 & 52.13133 & 0.67895 & 0.49830 & 0.39453 & 0.32332 \\
ru & 65039 & 51.75020 & 0.71578 & 0.54827 & 0.44681 & 0.37495 \\
zh & 56276 & 53.16430 & 0.64737 & 0.45399 & 0.34408 & 0.27072 \\