Datasets:

bigcode

/

the-stack-v2-train-full-ids

like 25

Hi @shawn0wang ! the-stack-v2-train-full-ids is the filtered version of the-stack-v2-dedup.
Also some language stats to compare the three versions: https://huggingface.co/datasets/bigcode/the-stack-v2/blob/main/language_stats.csv

Hi @anton-l , could you please clarify more on what filtering steps are applied to the-stack-v2-dedup and the-stack-v2-train-full-ids? Especially for the ones mentioned in the paper:

1. GitHub repositories, main branch, latest commits
2. Removed all files with >n (e.g., 100k, 512 ..) lines and longest line length >100k characters
3. Removed files classified as auto-generated
4. Removed files with <25% of alphabetic characters
5. Encoded data filter
6. Dedup, malicious code, opt-out ...

Thank you a lot!

Hi @Yarina ! The pipeline is roughly as follows:

• GitHub repositories, main branch, latest commits --> the-stack-v2
• Near-deduplication of the-stack-v2 --> the-stack-v2-dedup
• And the rest of the filters are applied to the-stack-v2-dedup to get the-stack-v2-train-full

I hope this helps!

Hi @anton-l , may I ask what is the storage size of the-stack-v2-train-full?

I see in README that the-stack-v2 is 67.5TB and the-stack-v2-dedup is 32.1TB, however I can't tell the storage size of the-stack-v2-train-full.

Ok from https://huggingface.co/datasets/bigcode/the-stack-v2/blob/main/language_stats.csv, the-stack-v2-train-full is roughly 2.39TB.