Guidelines for Machine Learning Artifact Submission

v20260724

icml-artifact-evaluation

Provides comprehensive guidelines for packaging supplementary materials, including code, data, model weights, and benchmarks, for major machine learning conferences like ICML. It details the requirements for both the double-blind review stage (focusing on anonymization and minimal replication commands) and the final public release package (ensuring permanent public links and clear licensing).

ICML Machine Learning Artifact Reproducibility Code Data Review Submission

Get Skill

338 downloads

Overview

ICML Artifact Evaluation

ICML does not let artifacts sit outside the paper's scientific argument. Reproducibility and code availability are explicitly considered in decision-making, and accepted submissions may publish the original supplementary material on OpenReview.

Review-stage package

Decide whether the artifact is code, data, model weights, simulator, benchmark, proof script, notebook, or supplementary manuscript.
Anonymize authors, repository ownership, filenames, commit history, logs, model cards, dataset cards, licenses, and personal paths.
If using an anonymous repository, put it on a branch that will not change after the submission deadline.
Put critical evaluation material in the paper body, not only in supplement. Reviewers decide whether to consult appendices or supplementary material.
Provide minimal commands, environment details, expected runtime, hardware assumptions, and result mapping.

Public-release package

Because accepted original supplementary material may become public, review-stage artifacts should not contain private, illegal, or unreleasable content.
For camera-ready, final supplementary material is not uploaded separately; code/data should move to a public repository or archive and be linked in the paper/OpenReview code URL field.
Add clear licenses and persistent identifiers when possible.

Anonymity leak checklist

ICML double-blind review means a single deanonymizing artifact can trigger a desk reject, so audit the package the way an adversarial reviewer would.

Leak vector	Where it hides	Mitigation
Repo ownership	Anonymous-repo account name, commit author	Use a fresh anonymized host, strip git history
File metadata	PDF author field, notebook kernel, model card	Clear metadata, rename author paths
Hard-coded paths	Cluster usernames in scripts and logs	Replace with placeholders before zipping
External links	Personal site, non-anonymous URL, shortener	Remove or route through an anonymous mirror

Worked vignette: optimizer artifact package

A paper shipping an adaptive optimizer includes training scripts, a pretrained checkpoint, and a proof-checking notebook. The review package gives minimal commands, expected runtime, and the hardware assumption so a reviewer can map a command to a benchmark number, while the checkpoint filename and the notebook kernel are scrubbed of the lab name. Because accepted ICML supplements can become public, the team confirms the checkpoint is releasable and the license is stated before the deadline, then plans the public repository and OpenReview code URL for camera-ready.

Output format

[Artifact role] code / data / model / benchmark / proof / none
[Review package] sufficient / incomplete / unsafe
[Anonymity risks] <metadata, repo, filenames, links>
[Decision relevance] <why reviewers need it>
[Public release plan] <repo/archive/license/code URL>

Info

Category Development

Name icml-artifact-evaluation

Version v20260724

Size 3.35KB

Source brycewang-stanford/Awesome-Journal-Skills

Updated At 2026-07-28