dataset_info:
features:
- name: instance_id
dtype: string
- name: version
dtype: string
- name: gold_patches
struct:
- name: code
dtype: string
- name: test
dtype: string
- name: test_patch
dtype: 'null'
- name: pre_patches
struct:
- name: code
dtype: string
- name: test
dtype: string
- name: pre_scripts
dtype: 'null'
- name: repo
dtype: string
- name: base_commit
dtype: string
- name: base_commit_timestamp
dtype: string
- name: hints_text
dtype: 'null'
- name: created_at
dtype: 'null'
- name: problem_statement
struct:
- name: code
dtype: string
- name: test
dtype: string
- name: environment_setup_commit
dtype: string
- name: evaluation
struct:
- name: FAIL_TO_PASS
sequence: string
- name: PASS_TO_PASS
dtype: 'null'
splits:
- name: test
num_bytes: 75055139
num_examples: 200
download_size: 20308767
dataset_size: 75055139
configs:
- config_name: default
data_files:
- split: test
path: data/test-*
Can Language Models Replace Programmers? REPOCOD Says 'Not Yet'
Large language models (LLMs) have achieved high accuracy, i.e., more than 90 pass@1, in solving Python coding problems in HumanEval and MBPP. Thus, a natural question is, whether LLMs achieve comparable code completion performance compared to human developers? Unfortunately, one cannot answer this question using existing manual crafted or simple (e.g., single-line) code generation benchmarks, since such tasks fail to represent real-world software development tasks. In addition, existing benchmarks often use poor code correctness metrics, providing misleading conclusions.
To address these challenges, we create REPOCOD, a code generation benchmark with 980 problems collected from 11 popular real-world projects, with more than 58% of them requiring file-level or repository-level context information. In addition, REPOCOD has the longest average canonical solution length (331.6 tokens) and the highest average cyclomatic complexity (9.00) compared to existing benchmarks. Each task in REPOCOD includes 313.5 developer-written test cases on average for better correctness evaluation. In our evaluations on ten LLMs, none of the models achieves more than 30 pass@1 on REPOCOD, disclosing the necessity of building stronger LLMs that can help developers in real-world software development.
For easier evaluation, we sample 200 of the hardest problems in REPOCOD to create REPOCOD-Lite, using the product of the prompt length and canonical solution length (in terms of line count) as an indicator of difficulty. From the three categories of questions—self-contained, file-level, and repo-level—we select 66, 67, and 67 samples respectively in descending order of the scores.
REPOCOD_Lite_Unified is a variation of REPOCOD-Lite that has a similar format as SWE-Bench for easier integration into the established inference pipelines.
For more details on data collection and evaluation results, please refer to our arxiv preprint.
Examples code for downloading repositories, preparing repository snapshot, and running test cases for evaluation are propived at code
Check our Leaderboard for preliminary results using SOTA LLMs with RAG.
"instance_id": Instance ID in REPOCOD
"version": Version of REPOCOD
"gold_patches": {
"code": Patch file to restore the target code,
"test": Patch file to restore the relevant tests for the target code
}
"test_patch": None,
"pre_patches": {
"code": Patch file to remove the target code,
"test": Patch file to remove the relevant tests for the target code
}
"pre_scripts": None,
"repo": {GitHub User Name}/{Project Name}
"base_commit": base commit
"base_commit_timestamp": time of the base commit
"hints_text": None,
"created_at": None,
"problem_statement": {
"code": Problem statement for code generation.
"test": Problem statement for test generation.
}
# "problem_statement_source": "repocod",
"environment_setup_commit": base commit
"evaluation": {
"FAIL_TO_PASS": list of relevant test cases
"PASS_TO_PASS": None, (all remaining tests that passes, we choose not to run the PASS_TO_PASS tests to avoid the computational cost)
}