SWE-Atlas-QnA / README.md
mohit-raghavendra's picture
Update README.md
98291bc verified
metadata
dataset_info:
  config_name: default
  splits:
    - name: test
      num_examples: 124

SWE-Atlas QnA

Codebase QnA is the first benchmark in the SWE-Atlas suite. It evaluates AI agents on deep code comprehension — tracing execution paths, explaining architectural decisions, and answering deeply technical questions about production-grade software systems.

124 tasks across 11 open-source repositories spanning Go, Python, C, and TypeScript.

Link to leaderboard - https://scale.com/leaderboard/sweatlas-qna

Schema

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 repo
repository_base_commit string 40-char commit SHA the environment is pinned to
language string go, python, c, or ts
category string Task category (see below)
rubric string (JSON) Evaluation criteria (see below)
docker_image string Docker Hub image for the sandboxed environment

Rubric format

Each task's rubric field is a JSON array:

[
  {
    "id": "a33fc01cba19849aaf3b55e6b801001c",
    "title": "1.1: States that kitty uses Unix sockets for external connections...",
    "annotations": {
      "type": "positive hli verifier",
      "importance": "must have"
    }
  }
]
  • positive hli verifier — a factual claim the answer must contain. If the claim is met my the agent's answer, the rubric item result is a PASS.
  • negative hli verifier — something the answer must not claim. If the claim is met my the agent's answer, the rubric item result is a FAIL.

Each task includes a docker_image field pointing to a pre-built Docker Hub image with the repository and all dependencies installed at /app:

Inference and Eval

We follow the standard SWE-Agent scaffold, and we provide a sample config (with the prompts) in default_qa_config.yaml

To run tasks, you can pull the docker image and run the container, and reset the environment to the base commit:

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

Evaluation is performed by an LLM judge (Claude Opus 4.5) that scores the agent's answer against each rubric criterion independently. Each criterion receives a binary score (met or not met) indicating and is then aggregated.

The primary metric is the Task Resolve Rate: the percentage of tasks for which the agent's answer is comprehensive (i.e. passes all rubric items and scores 1.0), as graded by a set of task-specific rubrics.

The agents are also instructed to avoid modifying source-code files, and clean up any temporary scripts made. So we add a programmatic check that fails a task that has any code changes after submission.

Our rubric evaluation prompt and other relevant details are in rubric_evaluation_config.yaml