README.md

metadata

license: mit
tags:
  - code-generation
  - code-completion
  - benchmark
  - software-engineering
pretty_name: DevBench

DevBench

This dataset packages the DevBench code-completion benchmark published by Microsoft, reformatted into a single data.jsonl file with a category field identifying the task category of each record.

Source

• Official repository: microsoft/devbench (benchmark/ folder)
• Retrieved: 2026-09-17, from the main branch of the repository (shallow clone).

Paper

Kumarappan, A., Golnari, P. A., Wen, W., Liu, X., Ryan, G., Sun, Y., Fu, S., & Nallipogu, E. (2026). DevBench: A Realistic, Developer-Informed Benchmark for Code Generation Models. arXiv:2601.11895. https://arxiv.org/abs/2601.11895

@misc{devbench2026,
  author        = {Kumarappan, Adarsh and Golnari, Pareesa Ameneh and Wen, Wen and Liu, Xiaoyu and Ryan, Gabriel and Sun, Yuting and Fu, Shengyu and Nallipogu, Elsie},
  title         = {{DevBench}: A Realistic, Developer-Informed Benchmark for Code Generation Models},
  year          = {2026},
  eprint        = {2601.11895},
  archivePrefix = {arXiv},
  primaryClass  = {cs.LG},
  url           = {https://arxiv.org/abs/2601.11895}
}

License

MIT License — Copyright (c) Microsoft Corporation, as stated in the LICENSE file of the source repository. Redistribute in accordance with the MIT License terms (retain the copyright notice).

File

• data.jsonl — 1,800 records, one JSON object per line, UTF-8, each with language and category fields.

Structure

DevBench is organized as benchmark/{language}/{category}/{category}.jsonl, i.e. 6 programming languages × 6 task categories × 50 tasks = 1,800 tasks. All 36 source shards were concatenated, each record tagged with:

• category — the task category, taken from the shard's parent directory name: api_usage, code2NL_NL2code, code_purpose_understanding, low_context, pattern_matching, syntax_completion (50 records each × 6 languages = 300 records per category).
• language — already present on every source record (python, javascript, typescript, java, cpp, c_sharp); kept as-is (re-added defensively only if a record were ever missing it, which did not occur here).

Each record also retains its original fields: id, testsource, prefix (code visible to the model before the cursor), golden_completion (reference answer), suffix (code visible after the cursor), and assertions (hidden unit-test-style checks used to grade a completion, not shown to the model under evaluation).

category	Records	Description
api_usage	300	Completions exercising a specific library/API call
code2NL_NL2code	300	Code↔natural-language translation tasks
code_purpose_understanding	300	Completions requiring understanding of surrounding code intent
low_context	300	Completions with minimal surrounding context
pattern_matching	300	Completions following a repeated code pattern
syntax_completion	300	Syntax-level fill-in-the-middle completions
Total	1,800

Known issues / decisions made while preparing this package

1. Only the benchmark/ folder (the actual task set) was packaged, per the task scope. The repository also ships completions/ (pre-generated model outputs from 9 models), judge_completions/ (LLM-judge scores), prompts/, evaluation/, and analysis/ — these are evaluation artifacts/tooling, not the benchmark data itself, and were left out of data.jsonl.
2. Each shard's companion *_formatted.txt file (a human-readable rendering of the same JSONL data, used for manual inspection) was skipped as redundant with the .jsonl source.
3. Record counts were verified as exactly 50 per shard × 36 shards = 1,800 before and after processing; no records were dropped or deduplicated.
4. Serialized with PowerShell's ConvertTo-Json (no Python available in the preparation environment); non-ASCII characters, if any, are escaped as \uXXXX, which is valid JSON and was validated line-by-line.