metadata
dataset_info:
features:
- name: id
dtype: string
- name: repo_url
dtype: string
- name: instruction
dtype: string
- name: parent_commit
dtype: string
- name: test_script
dtype: string
- name: testbed_environment
dtype: string
- name: requirements_txt
dtype: string
- name: solution_commit
dtype: string
- name: solution_patch
dtype: string
- name: modified_files
dtype: string
- name: language
dtype: string
splits:
- name: test
num_bytes: 4645427
num_examples: 100
configs:
- config_name: default
data_files:
- split: test
path: dataset.json
Dataset Summary
CoderEval is a codebase editing benchmark based on compact, natural language instructions. The task is to, given an edit instruction and a codebase, produce a patch file that makes the correct edit to the codebase.
The dataset was released as part of CoderEval: Evaluating the Capability of Language Models in Large-Scale Software Development
Dataset Structure
An example of a CoderEval task instance is as follows:
id (str) - A unique identifier for the task instance.
repo_url (str) - The URL of the repository involved in the task.
instruction (str) - The repository edit instruction.
parent_commit (str) - The commit hash of the repository representing the HEAD of the repository immediately before the instruction was carried out.
test_script (str) - The task's test suite as a python script to be run from the root of the repository.
testbed_environment (str) - The python version used to run the test suite.
requirements_txt (str) - The pip package dependencies required to run the test suite.
solution_commit (str) - The commit hash of the repository representing the HEAD of the repository immediately after the instruction was carried out.
solution_patch (str) - The patch in the unified diff format representing the difference between the parent and solution commit.
modified_files (str) - A dictionary of files and their content that were modified by the solution commit.
language (str) - The primary programming language of the repository.