README.md

metadata

tags:
  - code
size_categories:
  - 10M<n<100M

This is a filtered version of bigcode/the-stack-dedup that subsets for the following language splits:

[
  'ada', 'assembly', 'awk', 'c', 'cpp', 'c-sharp', 'clojure', 'common-lisp', 'dart',
  'erlang', 'f-sharp', 'fortran', 'glsl', 'go', 'haskell', 'isabelle', 'java', 'javascript',
  'julia', 'kotlin', 'llvm', 'markdown', 'mathematica', 'pascal', 'perl', 'php',
  'powershell', 'python', 'r', 'racket', 'ruby', 'rust', 'scala','shell', 'solidity',
  'sparql', 'sql', 'stan', 'typescript', 'zig'
]

Then, it has been filtered via the following steps:

1. For files forked more than 25 times, we retain them if the average line length is less than 120, the maximum line length is less than 300 and the alphanumeric fraction is more than 30%.
2. For files forked between 15 and 25 times, we retain them if the average line length is less than 90, the maximum line length is less than 150 and the alphanumeric fraction is more than 40%.
3. For files forked less than 15 times, we retain them if the average line length is less than 80, the maximum line length is less than 120 and the alphanumeric fraction is more than 45%.
4. The resultant language splits are then capped at 5 million samples.
5. Finally, an aggressive MinHash deduplication pipeline with a threshold of 0.5 and a shingle size of 20 is applied.