This repository is for the analysis done in BigCode Project. You can find analysis of datasets, models, architecture choices and more.
- Data analysis: In the folder
data_analysis, we analyze these two datasets: python-all-license (private) and python-safe-license. We provide the following statistics:- percentage of near duplicates
- percentage of configuration/test and uncommon files
- file size distribution
- loss analysis
- natural language distribution in comments/docstrings and number of files that can be successfully compiled
We also provide code to run near-deduplication, and to detect natural language of comments in Python datasets.
- Multi query attention experiments, for details refer here