mpasila's picture
Update README.md
4b65c80 verified
metadata
license: apache-2.0
task_categories:
  - translation
language:
  - ja
  - en

This is a modified version of NilanE/ParallelFiction-Ja_En-100k which has been turned into Alpaca format.

This has also been chunked for 4096 tokens for augmxnt/shisa-base-7b-v1 model's tokenizer.

If you want the non chunked version it's here.

Dataset format (correct one)

{
  'instruction' : 'Japanese chapter'
  'output' : 'English translation'
  'input' : 'empty'
}

Original Dataset card

Dataset details

Each entry in this dataset is a sentence-aligned Japanese web novel chapter and English fan translation. The intended use-case is for document translation tasks.

Dataset format

{
  'src' : 'JAPANESE CHAPTER'
  'trg' : 'ENGLISH TRANSLATION'
  'meta' : {
    "source": 'SAME ACROSS ALL ENTRIES',
    "series": 'NAME OF WEB NOVEL SERIES',
    "missed_lines": 'NUMBER OF LINES THAT WERE AT THE SAME INDEX BUT NOT DETECTED AS BEING TRANSLATIONS OF EACH OTHER',
    "inserted_lines_src":  'NUMBER OF LINES IN THE JAPANESE TEXT THAT DID NOT HAVE A MATCHING TRANSLATION BUT ARE BUFFERED BY TRANSLATED LINES',
    "inserted_lines_trg":  'SAME AS ABOVE BUT FOR ENGLISH',
    }
}

A high number of inserted lines is not necessarily a sign of a bad pair, as many translations concatenate or divide source chapters when publishing. Instead, watch out for high numbers of missed lines or entries where the inserted line count is high for both source and target.