Hugging Face's logo

Datasets:
cognitivedaniel
/
daniel-benchmark-v1

Tasks:
Text Generation
Modalities:
Tabular
Text
Formats:
json
Languages:
English
Size:
< 1K
Tags:
code
agent
swe-bench
bug-fixing
autonomous-agent
benchmark
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
1
daniel-benchmark-v1 / README.md
cognitivedaniel's picture
cognitivedaniel
Upload 2 files
93e865c verified
preview code
|
Raw History Blame Contribute Delete
3.19 kB
metadata 
license: cc-by-4.0
task_categories:
  - text-generation
  - text2text-generation
language:
  - en
tags:
  - code
  - agent
  - swe-bench
  - bug-fixing
  - autonomous-agent
  - benchmark
pretty_name: Daniel Benchmark v1 — Autonomous Agent Head-to-Head
size_categories:
  - n<1K

Daniel Benchmark v1

A public head-to-head benchmark comparing Daniel (Cognitive AI) and Devin (Cognition AI) on 6 real bugs from widely used open-source libraries.

Key results

  • Pass rate: Both agents 6/6 (100%)
  • Daniel cost: $1.45 total ($0.24/bug)
  • Devin cost: $4.91 total ($0.82/bug)
  • Daniel is 70% cheaper at equal accuracy

All pull requests are public. Graded by each project's own hidden test suite — neither agent saw the tests.

Dataset fields

Field Type Description
instance_id string Unique identifier for the bug instance
repo string owner/repo on GitHub
language string Primary language of the repo
issue_title string The bug description given to both agents
issue_url string Original GitHub issue URL
daniel_result string pass or fail
daniel_pr_url string Link to Daniel's pull request
daniel_time_s int Seconds from start to PR opened
daniel_cost_usd float Claude API cost in USD
devin_result string pass or fail
devin_time_s int Seconds from start to PR opened
devin_cost_usd float Devin session cost in USD
graded_by string Always maintainer_hidden_tests
benchmark_round int Iteration number (Daniel ran 4 rounds; table uses best passing round)

Repos tested

  • python-humanize/humanize — Python
  • more-itertools/more-itertools — Python
  • jshttp/mime-types — JavaScript (Express ecosystem)
  • sindresorhus/pretty-bytes — JavaScript
  • sindresorhus/camelcase — JavaScript
  • pillarjs/path-to-regexp — JavaScript (Express ecosystem)

Methodology

  1. Picked recent, real bug fixes from widely used libraries
  2. Forked each project at the commit just before the maintainers' own fix
  3. Gave both agents an identical issue-style prompt with no hints about the fix
  4. Graded each PR by overwriting test files with the maintainer versions and running the project's test command
  5. All PRs are public

See also

Upcoming: v2 (10-bug suite)

The next version will include 10 harder bugs from Django, FastAPI, SQLAlchemy, TypeScript, Next.js, Elasticsearch, Kubernetes, Spring Boot, Pandas, and Apache Airflow — testing multi-file and cross-language fixes. Competitors will include OpenHands and SWE-agent in addition to Devin.

Citation 

@misc{cognitive-ai-daniel-benchmark-2026,
  title={Daniel Benchmark v1: Autonomous AI Agent Head-to-Head on Real Open-Source Bugs},
  author={Cognitive AI},
  year={2026},
  url={https://huggingface.co/datasets/cognitivedaniel/daniel-benchmark-v1}
}