cliptrace-2026/cliptrace-baseline-data

CLIPTrace 2026 Baseline Data

Participant data for the reproducible CLIPTrace 2026 baseline. The repository mirrors the full-size Imagenette train and validation images used by the baseline and preserves the original class-directory layout.

This repository is intended to be public with access gating. Before downloading, users must accept the repository terms and the upstream image-use conditions configured by the organizers on the Hugging Face settings page.

Dataset summary

Split	Images	Classes
train	9,469	10
validation	3,925	10
total	13,394	10

Expected layout:

imagenet/
  train/
    n01440764/
    ...
  val/
    n01440764/
    ...
manifest/
  files.sha256
  dataset-summary.json

The baseline uses a seeded, recorded subset of imagenet/val. Labels are not used by the trigger-inversion objective but the class folders are preserved for provenance and compatibility with ImageFolder.

Source and rights

Imagenette was prepared by fast.ai as a small, easily usable subset of ImageNet. See the upstream Imagenette repository for its description and download information.

The images originate from ImageNet. The dataset repository does not grant new rights in the images. Users are responsible for complying with the ImageNet terms of access and use, research-only restrictions where applicable, and any rights associated with individual images. Access may be revoked if the data is redistributed or used outside the accepted terms.

Integrity and reproducibility

manifest/files.sha256 records the SHA-256 digest of every image using POSIX relative paths. manifest/dataset-summary.json records file counts by split and class. Generate both with the baseline repository's tools/build_dataset_manifest.py; verify them again after Hub download before running a formal experiment.

Personal and sensitive information

Imagenette includes ordinary photographs selected from ImageNet. Some images may incidentally contain people. The organizers have not performed identity or consent annotation. Do not use this mirror for face recognition, identity inference, surveillance, or other decisions about individuals.

Citation

Please cite the upstream Imagenette project and ImageNet when using these images, and cite the CLIPTrace 2026 baseline repository when reporting results obtained with its deterministic sampling protocol.