Description

YT_subtitles is a tool for building a dataset from youtube subtitles. It extracts the (non machine-generated) subtitles from all the videos returned by a list of search terms.

The resulting files contain a string of text per language, per minute of subtitles, with the name of the language appended as a header. The dataset is designed to improve the multilingual performance of language models trained on it. If only a single language is available, the output is just a text version of the subtitles, with no metadata.

Download a pre-processed dataset scraped using the search terms in examples.csv here

The dataset is a jsonl file compressed with zstd. You can easily read the files using lm-dataformat.

Setup

pip install -r requirements.txt

If you're running this on an Ubuntu server, I've found I need to install the following packages to get requests-html working

sudo apt-get install gconf-service libasound2 libatk1.0-0 libatk-bridge2.0-0 libc6 libcairo2 libcups2 libdbus-1-3 libexpat1 libfontconfig1 libgcc1 libgconf-2-4 libgdk-pixbuf2.0-0 libglib2.0-0 libgtk-3-0 libnspr4 libpango-1.0-0 libpangocairo-1.0-0 libstdc++6 libx11-6 libx11-xcb1 libxcb1 libxcomposite1 libxcursor1 libxdamage1 libxext6 libxfixes3 libxi6 libxrandr2 libxrender1 libxss1 libxtst6 ca-certificates fonts-liberation libappindicator1 libnss3 lsb-release xdg-utils wget

Usage

CLI for YT_subtitles - extracts Youtube subtitles from a list of search terms

positional arguments:
  search_terms         A comma separated list of search terms, alternatively,
                       pass path to a csv with the -c argument

optional arguments:
  -h, --help           show this help message and exit
  --out_path OUT_PATH  Output location for final .txt files (default "output")
  -s, --save_links     whether to save links to a .csv file (default True)
  -c, --csv            if true, positional arg should be a path to a .csv file
                       containing search terms (default False)
  --scroll SCROLL      how far to scroll down in the youtube search (default 1000)

e.g

python get_subs.py -c examples.csv

or

python get_subs.py "movie review,GPT-3,true crime documentary"

TODO:

• output as lm dataformat