Manually Proofread ImageNet-Aligned Benchmark for OOD Evaluation

reLAIONet is a manually proofread, web-sourced image classification benchmark aligned to ImageNet's label space. It is designed for out-of-distribution evaluation of class-conditional generative and discriminative models. The dataset was created by harvardairobotics and was last updated in April 2026.

Use Cases

• Evaluate out-of-distribution performance of classifiers using the ImageNet-aligned labels.
• Benchmark class-conditional generative models against the manually proofread image-label pairs.
• Assess model robustness on web-sourced images distinct from standard validation sets like ImageNet val.

Strengths

• Manually proofread to ensure label accuracy.
• Aligned to the established ImageNet label space for direct comparison.

Limitations

• Specific sample size and class distribution details are not provided.
• The web-sourced nature may introduce unknown biases compared to curated datasets.

Provenance

Source

harvardairobotics via Hugging Face.

Collection Method

Manually proofread and web-sourced.

Freshness

Last updated in April 2026.