SWE-Atlas-QnA-AA

SWE-Atlas-QnA-AA is Scale AI's SWE-Atlas-QnA as graded in the Artificial Analysis Coding Agent Index. The task set is the same as upstream (same questions, repos and pinned commits). The grading changes are there to improve the stability and consistency of verdicts.

Scale's paper: SWE Atlas: Benchmarking Coding Agents Beyond Issue Resolution. Code and task packages: github.com/scaleapi/SWE-Atlas.

Unchanged from upstream

• The 124 tasks: questions, repositories, pinned base commits, Docker images and reference answers.
• The 1,306 rubric criteria: ids, order, polarity and importance.
• Scoring: a task is resolved only if every criterion is met (Task Resolve Rate). No partial credit.

Grading methodology

• Judge: GPT-6 Luna (xhigh) instead of Claude Opus 4.5.
• 5 votes per criterion, majority vote wins. A negative criterion's polarity is applied after the vote.
• Structured outputs: the judge replies in a fixed JSON format (reason, evidence, verdict), giving its reasoning and supporting quotes from the answer before the YES/NO verdict.
• Answer date in the prompt, so "latest"/"fresh" are judged as of the run.
• Observed results: where a criterion asks for observed or measured results, results inferred from reading the code don't count; results the answer reports from running something do.

The full configuration (model, voting, output schema, prompts) is in rubric_evaluation_config.yaml. Scale's original is the previous version of that file in the commit history.

Criteria

127 of 1,306 reworded or adjusted:

• Stability (75): clearer wording where judges often split votes or read the criteria differently.
• Precision (52): small tweaks so the wording better aligns with the code at the pinned commit.

These changes came from several rounds of audits on our rollouts and re-testing against the code and real answers. We kept the original wording where the evidence wasn't clear-cut.

This dataset also includes a fix from Scale's GitHub repo: 9acc18e ("Fix inverted QnA rubric polarity", 16 July 2026) corrects two negative criteria in the grafana/k6 task (6905333b74f22949d97baa22) that were phrased as "Avoids ...", which inverts their meaning under the negative-criterion rule. These two items are marked scale-correction.

Each rubric item records Scale's text and the kind of change (below).

Schema

Columns are Scale's, unchanged.

Column	Type	Description
task_id	string	Unique 24-char hex identifier
prompt	string	The question presented to the agent
reference_answer	string	Expert-written reference answer
repository_url	string	GitHub repository (owner/name)
repository_base_commit	string	40-char commit SHA the environment is pinned to
language	string	go, python, c or ts
category	string	Task category
rubric	string (JSON)	Evaluation criteria (see below)
docker_image	string	Docker image for the sandboxed environment, with the repository at /app

Rubric format

rubric is a JSON array. Each item keeps Scale's id, title and annotations and adds original_title and change:

{
  "id": "4ca20ef4bbf52c6414a7478d688d08c3",
  "title": "1.7: States that kitty sets environment variables KITTY_PID, KITTY_LISTEN_ON, and KITTY_WINDOW_ID for child processes.",
  "annotations": {
    "type": "positive hli verifier",
    "importance": "must have"
  },
  "original_title": "1.7: States that shell integration works through environment variables KITTY_PID, KITTY_LISTEN_ON, and KITTY_WINDOW_ID that kitty sets for child processes.",
  "change": "precision"
}

• title: the criterion text we grade (the judge gets it without the 1.7: numbering).
• original_title: Scale's text. Equal to title when unchanged.
• change: null, stability, precision or scale-correction.
• annotations.type: positive hli verifier (the answer must make the claim) or negative hli verifier (it must not). A negative criterion is met when the judge's verdict is NO.
• annotations.importance: always must have.

Running

Each docker_image holds the repository and its dependencies at /app. Reset to the pinned commit before the agent starts:

cd /app
git config --global --add safe.directory /app
git restore .
git reset --hard <repository_base_commit>
git clean -fdq

We run each coding agent through its own harness on the task prompt and grade its written answer with the judge above. See the Coding Agent Index methodology. Scale's SWE-Agent config, default_qa_config.yaml, is kept for reference; we don't use it.

Files

• data/test-00000-of-00001.parquet: the 124 tasks with the SWE-Atlas-QnA-AA rubric text.
• rubric_evaluation_config.yaml: judge configuration.
• default_qa_config.yaml: Scale's SWE-Agent config, unchanged.
• LICENSE, NOTICE.

Citation

SWE-Atlas-QnA is by Scale AI. Please cite their paper, and credit Artificial Analysis for the SWE-Atlas-QnA-AA grading.

@misc{raghavendra2026sweatlasbenchmarkingcoding,
      title={SWE Atlas: Benchmarking Coding Agents Beyond Issue Resolution},
      author={Mohit Raghavendra and Soham Dan and Miguel Romero Calvo and Yannis Yiming He and Johannes Baptist Mols and Gautam Anand and Cole McCollum and Edgar Arakelyan and Vijay Bharadwaj and Andrew Park and Jeff Da and MohammadHossein Rezaei and Bing Liu and Brad Kenstler and Yunzhong He},
      year={2026},
      eprint={2605.08366},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2605.08366},
}

License

Apache 2.0, same as upstream. See LICENSE and NOTICE.