README.md

metadata

license: other
pretty_name: Token-CX Artifacts
tags:
  - explainable-ai
  - computer-vision
  - vision-transformer
  - concept-based-explanations
  - imagenet
  - datasets
  - tabular
configs:
  - config_name: bank_samples
    data_files:
      - split: train
        path: manifests/bank_samples.parquet
    default: true
  - config_name: evaluation_samples
    data_files:
      - split: validation
        path: manifests/evaluation_samples.parquet
  - config_name: deit_exemplars
    data_files:
      - split: train
        path: banks/deit_b16_distilled/exemplars.parquet
  - config_name: deit_diagnostics
    data_files:
      - split: train
        path: banks/deit_b16_distilled/diagnostics.parquet
  - config_name: vit_exemplars
    data_files:
      - split: train
        path: banks/vit_b16/exemplars.parquet
  - config_name: vit_diagnostics
    data_files:
      - split: train
        path: banks/vit_b16/diagnostics.parquet

Token-CX Artifacts

This dataset repository contains the numerical class-specific concept banks and reproducibility metadata used by Token-CX: Token Concept-Based Explanations for Vision Transformers.

Contents

• banks/*/bases.safetensors: fixed class-specific NMF bases for 1,000 ImageNet classes.
• banks/*/exemplars.parquet: identifiers and patch locations of the top-activating training examples for each concept.
• banks/*/diagnostics.parquet: per-class NMF diagnostics.
• manifests/*: the exact training and evaluation manifests.
• artifact_manifest.json and checksums/sha256sums.txt: provenance and file integrity metadata.

The released banks use Transformer Block 9 and concept rank K=4. Each class bank was constructed from 100 ImageNet training images while retaining all 196 patch locations per image. ViT-B/16 and DeiT-B/16-Distilled use the same training and evaluation manifests.

ImageNet notice

This repository does not redistribute ImageNet images or image crops. Manifest and exemplar rows contain identifiers and relative paths only. Users must obtain ImageNet separately and comply with the ImageNet terms of access.

Code

The implementation is maintained at https://github.com/VuTrinhNguyenHoang/Token-CX.