README.md

metadata

license: apache-2.0
dataset_info:
  features:
    - name: repo
      dtype: string
    - name: commit_hash
      dtype: string
    - name: completion_file
      struct:
        - name: filename
          dtype: string
        - name: content
          dtype: string
    - name: completion_lines
      struct:
        - name: infile
          sequence: int32
        - name: inproject
          sequence: int32
        - name: common
          sequence: int32
        - name: commited
          sequence: int32
        - name: non_informative
          sequence: int32
        - name: random
          sequence: int32
    - name: repo_snapshot
      sequence:
        - name: filename
          dtype: string
        - name: content
          dtype: string
    - name: completion_lines_raw
      struct:
        - name: commited
          sequence: int64
        - name: common
          sequence: int64
        - name: infile
          sequence: int64
        - name: inproject
          sequence: int64
        - name: non_informative
          sequence: int64
        - name: other
          sequence: int64
configs:
  - config_name: large_context
    data_files:
      - split: test
        path: data/large_context/*
  - config_name: medium_context
    data_files:
      - split: test
        path: data/medium_context/*
  - config_name: small_context
    data_files:
      - split: test
        path: data/small_context/*
  - config_name: huge_context
    data_files:
      - split: test
        path: data/huge_context/*

LCA Project Level Code Completion

How to load the dataset

from datasets import load_dataset

ds = load_dataset('JetBrains-Research/lca-codegen-huge', split='test')

Data Point Structure

• repo – repository name in format {GitHub_user_name}__{repository_name}
• commit_hash – commit hash
• completion_file – dictionary with the completion file content in the following format:
• filename – filepath to the completion file
• content – content of the completion file
• completion_lines – dictionary where keys are classes of lines and values are a list of integers (numbers of lines to complete). The classes are:
• committed – line contains at least one function or class that was declared in the committed files from commit_hash
• inproject – line contains at least one function or class that was declared in the project (excluding previous)
• infile – line contains at least one function or class that was declared in the completion file (excluding previous)
• common – line contains at least one function or class that was classified to be common, e.g., main, get, etc (excluding previous)
• non_informative – line that was classified to be non-informative, e.g. too short, contains comments, etc
• random – randomly sampled from the rest of the lines
• repo_snapshot – dictionary with a snapshot of the repository before the commit. Has the same structure as completion_file, but filenames and contents are orginized as lists.
• completion_lines_raw – the same as completion_lines, but before sampling.

How we collected the data

To collect the data, we cloned repositories from GitHub where the main language is Python. The completion file for each data point is a .py file that was added to the repository in a commit. The state of the repository before this commit is the repo snapshot.

Huge dataset is defined by number of characters in .py files from the repository snapshot. This number larger then 768K.

Dataset Stats

• Number of datapoints: 296
• Number of repositories: 75
• Number of commits: 252

Completion File

• Number of lines, median: 313.5
• Number of lines, min: 200
• Number of lines, max: 1877

Repository Snapshot

• .py files: median 261, from 47 to 5227
• non .py files: median 262, from 24 to 7687
• .py lines: median 49811
• non .py lines: median 60163

Line Counts:

• infile: 2608
• inproject: 2901
• common: 692
• committed: 1019
• non-informative: 1164
• random: 1426
• total: 9810

Scores

HF Space