README.md

metadata

license: apache-2.0
language:
  - en
tags:
  - lean
  - openai
  - math
  - clustering
  - code
size_categories:
  - 100K<n<1M

Dataset for OpenAI Lean Code grouping

Published Lean code is clustered via combinatorial hierarchical clustering

This is graph based combinatorial clustering not a statistical clustering in k-means clustering. 

This combinatorial clustering finds context and meaning among the data. 

Details

• Total Files in the Github repo is 122141
• Total Lean code lines 22358566
• Total Lines taken for Clustering 5000000
• Total Documents taken for Clustering 25391
• Number of Code Line Clusters generated 248289
• Number of File Clusters generated 3131
• Total Lines in these clusters 3628607
• Total Docs in these clusters 24738
• % of lines which are clustered 72%
• % of docs which are clustered 97%

How to understand the data

• Finding similar contextual blocks allows the Lean code to be understood easily. Because the amount of generated code is so large that entire proofs can't be manually evaluated. 

• Finding similarity allows for easier code understanding because contextual comparison links both different part of code, files and domins easily. 

• Since this is LLM-generated Lean code for the mathematical proof, it will make tracing the reasoning or the method of LLM generation possible. 

• Full contextual clustering will be updated soon.

 

Method

1. The Lean code data is combined from all Lean code files in the GitHub repo.
2. The code data is fed to the Graph AI based combinatorial clustering method directly. No change.
3. The Graph AI compares the code lines via combinatorial clustering.
4. Resulting clusters of code lines.

Uses

• Can be used understand how the Lean code is generated by LLMs.

Schema

All files are new line JSON files. It could be loaded as JSONL

1. data_file.json - Base Data file

• doc_id - Unique ID for the file.
• data - data from lean code files.
• folder_name - folder name in which the Lean code file is located.
• domain_name - domain name from the folder name location.
• file_name - Lean code file name.

2. line_data.json - Contains lines used for Clustering

• doc_id - Unique ID
• line_id - Unique ID
• line - Actual Lean Code line as it is.

3. rows_clusters.json

• cluster_id - Unique ID denoting cluster.
• cluster - List of line_id that forms a conceptual cluster.
• confidence - Score of confidence showing how much these line_id are related.

4. docs_clusters.json

• cluster_id - Unique ID denoting cluster
• cluster - List of doc_id that forms a conceptual cluster.
• confidence - Score of confidence showing how much these doc_id are related.