Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →brycewang-stanford/cvpr-artifact-evaluation
Use when packaging code, models, datasets, or demo videos for a CVPR paper at either review time or release time, covering anonymous supplement packaging under the external-link ban, the dataset-release-by-camera-ready rule, model-weight and license decisions, and making a vision artifact runnable by a skeptical stranger.
Passed 19 of 22 test cases with the skill, and 11 of 22 without it.
and 6 more in the full report
8 fixed11 passed either way3 still failing0 broke
Tokens, turns and time describe resource use, not a dollar price.
One of 8 test cases the skill fixed. Same model, same prompt, with and without the skill.
We just got accepted to CVPR and are setting up our repository for public release alongside our camera-ready paper. We have model checkpoints, a dataset subset, and training scripts. Please evaluate our package and organize your feedback into six sections: Artifact stage, Identity scan (clean or hit list), Runnability details (entrypoint, pins, mapping to paper outputs), Weights & data plan, Dataset clause compliance, and an ordered list of Gaps to address.
| Checked | Without skill | With skill |
|---|---|---|
| The output specifies that dataset contributions must be made publicly available by the camera-ready deadline. | Missed | Passed |
| The output recommends tagging the repository with a version tag at camera-ready. | Missed | Passed |
The instructions your model reads when the skill loads.
CVPR has no badge-granting artifact committee; the "evaluation" of your artifacts is done informally, twice, by two different audiences. At review time, reviewers poke at an anonymous supplement while deciding whether to trust your tables. After acceptance, the entire field — 2026 drew 44,011 authors alone — decides whether your method becomes a baseline or a citation nobody reproduces. Package for both audiences deliberately.
Everything must ride inside the uploaded supplement, because the 2026-verified policy bans external links that expand content or subvert review. That kills the usual anonymous-repo-link workflow many venues tolerate. Practical consequences:
.git history, no usernames in
paths, no institutional cluster hostnames, no wandb entity names in configs.cvpr-supplementary) but strip
container metadata — video files carry author tags more often than PDFs do.# Build a clean review artifact from a working repo
git archive --format=tar HEAD | tar -x -C /tmp/artifact # no .git, no untracked junk
cd /tmp/artifact
grep -rniE "$(whoami)|<lab-name>|<cluster-host>|wandb\.(ai|entity)" . | head # identity scan
pip-compile --quiet requirements.in # pin, don't approximate
zip -r ../supplement_code.zip . -x '*.ckpt' '*.pth' # weights out, recipe in
A reviewer who opens your archive gives it minutes, not an afternoon. One entry point, one config, one expected number:
| Artifact layer | Minimum viable | Gold standard |
|---|---|---|
| Environment | pinned requirements.txt + CUDA/driver note | Dockerfile reproducing the CRF hardware row |
test cases passed, vs 11 / 22 without
Published skills whose descriptions closely match this one, with their headline benchmark lift.
More skills from the same source and nearby in the registry, with their headline benchmark lift.