Hugging Face's logo

Datasets:
lang-uk
/
every_prompt

Tasks:
Question Answering
Modalities:
Tabular
Text
Size:
1M - 10M
Libraries:
Datasets
License:
Dataset card Viewer Files Community
every_prompt / README.md
Dmitry Chaplinsky
Another fix for README
dace661
preview code
|
raw
history blame
3.23 kB
metadata
license: mit
task_categories:
  - question-answering
pretty_name: Every Prompt
size_categories:
  - 1M<n<10M
multilinguality:
  - multilingual

Every Prompt

Every Prompt is a data-driven approach to mining instructions from the web. It contains over a million FAQs and HowTos from around the world in a structured format. It also has basic pre-processing to calculate the length of the useful text and identify the language of that text with the help of GCLD3

It relies on the Web Data Commons dataset (from October 2022) to find the seed list of sites with HowTo and FAQPage items. The general pipeline looks like this:

  • Download 1.6TB of structured data from webdatacommons to identify the pages with the structured data we need (wget/parallel). That gives us 1,985,925 seed pages
  • Crawls the seed pages and tries to extract structured data using extruct package. That left around 1,358,638 pages which are alive and well-formed.
  • Extracts only the relevant structured data of the HowTo/FAQPage type with the help of jmespath. That boils down to 1,266,926 json documents.
  • Extracts the textual information out of the structure to identify the text's language, the textual data's length, and the text/data ratio.

You can use the resulting dataset by filtering for the language and amount of the text. You need to convert the structured data into instructions yourself. You'll need to apply extra cleansing/evaluation of the instructions you've got because, you know, the internet is still full of crap.

Caveat emptor: the format of the FAQs and HowTo's in the dataset might vary greatly. Account for that. To understand potential pitfalls, look at the jmespath expression at the export_structured_data.py.

Recreating the results

  1. Clone the repo without the LFS files.
  2. Install requirements from requirements.txt.
  3. Install pv and parallel.
  4. Run bin/get_seed_urls.sh to filter urls of interest out of 1.6TB of compressed data. Don't worry about disk space. Worry about the traffic. That will take around 5h on decent connection.
  5. Run scrapy spider like this scrapy crawl webdatacommons_org -s WEB_DATA_COMMONS=web_data_commons_urls_sample.txt -L INFO -o webdatacommons.jsonlines with WEB_DATA_COMMONS pointing to the list of seed URLs from step 4. That might take up to a few weeks.
  6. Run python extract_relevant_structured_data.py --num-threads 12 webdatacommons.jsonlines relevant.jsonlines.bz2. That's fast, probably around 30 minutes.
  7. Run python export_structured_data.py relevant.jsonlines.bz2 extruct_out.jsonlines.bz2 to obtain the final version of the dataset.

Advices

If you want to recreate the results:

  • Get yourself a server or VPS with enough space (80GB should be enough).
  • Look at the code. You'd probably want to make changes here and there.
  • All the python scripts have extra parameters to control the number of threads and the chunk size. Both accept compressed input and output files with the help of smart_open lib.

License

Code of the project has an MIT license.