Data resource

Code-QA-Bench data and reproduction

Inspect real tasks, schema, and release versions. Download the data, run a model with the public evaluation framework, and generate your own results.

Public tasks
628
Task files
2
Data format
JSON
Reproduction status
Partially reproducible

Tasks and framework are public; paper run outputs and judge records are not.

01

Download data

Every file comes from the pinned repository commit. This page does not repackage or modify task content.

Code and task snapshot

Complete repository archive at pinned commit d615139.

ZIPabout 5 MB Open / download
Doc-dependent

The 100-task diagnostic set for documentation utility.

JSON400 KB Open / download
02

Task preview

These specimens come from the public task files and expose the input, answer scope, and verification rubric.

Task context
Task ID
code-derivable:sympy_gen_03
Repository
sympy @ 693a559a
Key files
sympy/functions/elementary/_trigonometric_special.py · trigonometric.py · sympy/core/intfunc.py
Question

Where does `ipartfrac` get called, and how does its data flow through `migcdex`, `igcdex`, and `gcdext`?

Verification rubric
  1. Identify that ipartfrac is called in cos._eval_rewrite_as_sqrt, not directly in the sin implementation.
  2. Explain that ipartfrac uses functools.reduce with a multiplication function to compute the denominator product.
  3. Explain that complement values are constructed with denom // x and passed to migcdex.
03

Dataset schema

The primary and diagnostic task sets share the same core JSON fields.

FieldTypeRequirementDescription
idstringRequired

Unique task identifier within its task set.

repostringRequired

Pinned source repository used by the task.

questionstringRequired

Repository-level code question presented to the evaluated model.

categoryenumRequired

Question category: what, why, where, or how.

sub_typestringRequired

More specific capability type such as architecture_exploration.

gold_answerstringRequired

Reference answer produced by the tool-equipped generation workflow.

rubricstring[]Required

Atomic criteria used by the judge to score a model answer.

key_filesstring[]Required

Repository paths supporting the reference answer.

source_docstringRequired

Original documentation excerpt used to establish the task topic.

verification_verdictenumRequired

Automated audit verdict: pass, warn, or fail.

verification_issuesstring[]Required

Issues and evidence boundaries found by automated verification.

strip_verify_leakagenumberOptional

Estimated leakage remaining after documentation stripping.

04

Reproduce the evaluation

Commands come from the repository README. Reruns require model and judge API credentials, and results may vary with endpoint versions.

  1. 01

    Install

    Create an environment and install the repository in editable mode.

    uv venv
    source .venv/bin/activate
    uv pip install -e ".[dev]"
  2. 02

    Prepare repositories

    Fetch pinned repositories from repos.json and build stripped variants.

    code-qa-bench setup-repos
    code-qa-bench strip-repos
  3. 03

    Run a model

    Run the primary task set in the Code-only condition; credentials live in config.json.

    code-qa-bench run --model my-model --condition stripped
  4. 04

    Judge and report

    Score saved answers with a judge and produce an aggregate report.

    code-qa-bench judge results/my-model.json --judge-model judge-model
    code-qa-bench report results/my-model_judged_by_judge-model.json
05

Evidence and versions

Separates inspectable data, paper-reported results, and run-level evidence that is not yet public.

01
Tasks and repositories

The repository includes 528 code-derivable and 100 doc-dependent tasks plus pinned commits for ten upstream repositories.

Publicly inspectable
02
Scores and intervals

Scores and 95% CIs come from arXiv:2605.29277v1; intervals use 10,000 bootstrap resamples with seed 42.

Paper-reported
03
Run-level evidence

The paper reports aggregate tables, but the reviewed public commit does not include task-level model outputs or judge results, so results are not marked fully verified.

Partially public
04
Human review

Version 1 reports neither human validation nor inter-judge agreement; both remain explicit limits on formal comparison.

Not performed
Current reproduction boundary

Tasks and aggregate results are publicly inspectable; per-task model outputs, judge records, and human validation are not yet fully public.