README.md

metadata

license: apache-2.0
pretty_name: Encoding Mismatch Analysis Data
tags:
  - computer-vision
  - vision-transformer
  - knowledge-distillation
  - representation-analysis
  - encoding-mismatch
  - spectral-energy-pattern
  - pca
  - svd
  - icml-2026
  - arxiv:2511.15572
configs:
  - config_name: npz_array_catalog
    data_files:
      - split: train
        path: data/npz_array_catalog.csv
    default: true
  - config_name: manifest
    data_files:
      - split: train
        path: data/manifest.csv
  - config_name: cait_sep_sep_thresholds
    data_files:
      - split: train
        path: raw/cait/sep/sep_thresholds.csv
  - config_name: comparison_sep_sep_comparison_table
    data_files:
      - split: train
        path: raw/comparison/sep/sep_comparison_table.csv
  - config_name: deit_small_sep_sep_thresholds
    data_files:
      - split: train
        path: raw/deit_small/sep/sep_thresholds.csv
  - config_name: swin_small_sep_sep_thresholds
    data_files:
      - split: train
        path: raw/swin_small/sep/sep_thresholds.csv
  - config_name: vit_base_patch14_dinov2_sep_sep_thresholds
    data_files:
      - split: train
        path: raw/vit_base_patch14_dinov2/sep/sep_thresholds.csv
  - config_name: vit_base_patch16_224_dino_sep_sep_thresholds
    data_files:
      - split: train
        path: raw/vit_base_patch16_224_dino/sep/sep_thresholds.csv
  - config_name: vit_base_patch16_224_mae_sep_sep_thresholds
    data_files:
      - split: train
        path: raw/vit_base_patch16_224_mae/sep/sep_thresholds.csv
  - config_name: vit_base_patch16_clip_openai_sep_sep_thresholds
    data_files:
      - split: train
        path: raw/vit_base_patch16_clip_openai/sep/sep_thresholds.csv
  - config_name: vit_huge_patch14_224_mae_sep_sep_thresholds
    data_files:
      - split: train
        path: raw/vit_huge_patch14_224_mae/sep/sep_thresholds.csv
  - config_name: vit_large_21k_in1k_sep_sep_thresholds
    data_files:
      - split: train
        path: raw/vit_large_21k_in1k/sep/sep_thresholds.csv
  - config_name: vit_large_patch14_clip_openai_sep_sep_thresholds
    data_files:
      - split: train
        path: raw/vit_large_patch14_clip_openai/sep/sep_thresholds.csv
  - config_name: vit_large_patch14_dinov2_sep_sep_thresholds
    data_files:
      - split: train
        path: raw/vit_large_patch14_dinov2/sep/sep_thresholds.csv
  - config_name: vit_large_patch16_224_mae_sep_sep_thresholds
    data_files:
      - split: train
        path: raw/vit_large_patch16_224_mae/sep/sep_thresholds.csv
  - config_name: vit_small_patch16_224_dino_sep_sep_thresholds
    data_files:
      - split: train
        path: raw/vit_small_patch16_224_dino/sep/sep_thresholds.csv
  - config_name: vit_tiny_patch16_224_21k_sep_sep_thresholds
    data_files:
      - split: train
        path: raw/vit_tiny_patch16_224_21k/sep/sep_thresholds.csv

Encoding Mismatch Analysis Data

This repository publishes the prepared numerical analysis artifacts associated with From Per-Image Low-Rank to Encoding Mismatch: Rethinking Feature Distillation in Vision Transformers. It is analysis data, not an image or model-training dataset, and it does not redistribute ImageNet.

Links

• Paper: https://arxiv.org/abs/2511.15572
• Hugging Face paper page: https://huggingface.co/papers/2511.15572
• Code and analysis scripts: https://github.com/thy960112/From-Per-Image-Low-Rank-to-Encoding-Mismatch
• Lift and WideLast checkpoints: https://huggingface.co/Huiyuancs/Encoding_Mismatch

Load the default configuration

The default npz_array_catalog configuration has one row per safely inspected array inside the original NPZ files. It records the source file, array key, dtype, JSON-encoded shape, dimensionality, element count, finite numeric summary statistics where applicable, a small JSON preview, and any safe inspection error.

from datasets import load_dataset

catalog = load_dataset(
    "Huiyuancs/Encoding_Mismatch_Analysis_Data",
    split="train",
)

Load the manifest

The manifest records the repository-relative path, file type, byte size, SHA-256 digest, and recommended loader for every artifact copied from the GitHub repository's Raw data/ directory.

from datasets import load_dataset

manifest = load_dataset(
    "Huiyuancs/Encoding_Mismatch_Analysis_Data",
    "manifest",
    split="train",
)

Load an original CSV table

Each original CSV has a separate configuration. For example:

from datasets import load_dataset

table = load_dataset(
    "Huiyuancs/Encoding_Mismatch_Analysis_Data",
    "cait_sep_sep_thresholds",
    split="train",
)

Download and read an original NPZ file

Use hf_hub_download for the original binary artifacts and keep NumPy's pickle loading disabled:

from huggingface_hub import hf_hub_download
import numpy as np

path = hf_hub_download(
    repo_id="Huiyuancs/Encoding_Mismatch_Analysis_Data",
    repo_type="dataset",
    filename="raw/cait/dataset_pca/dataset_pca_results.npz",
)

with np.load(path, allow_pickle=False) as archive:
    print(archive.files)

The same download method can be used with any relative_path from the manifest configuration.

Repository structure

README.md
data/
├── manifest.csv
├── npz_array_catalog.csv
└── viewer_csv/              # only created when a source CSV needs it
raw/                         # byte-identical copy of Raw data/

data/npz_array_catalog.csv is a compact inspection index, not a replacement for the original arrays. data/manifest.csv supplies checksums for verifying the originals. All released CSV files load directly with Hugging Face Datasets, so their configurations point to the byte-identical files under raw/; no viewer-normalized copies were needed.

CSV configurations

Configuration	Original file	Config data file	Rows	Representation
cait_sep_sep_thresholds	raw/cait/sep/sep_thresholds.csv	raw/cait/sep/sep_thresholds.csv	5	original
comparison_sep_sep_comparison_table	raw/comparison/sep/sep_comparison_table.csv	raw/comparison/sep/sep_comparison_table.csv	14	original
deit_small_sep_sep_thresholds	raw/deit_small/sep/sep_thresholds.csv	raw/deit_small/sep/sep_thresholds.csv	5	original
swin_small_sep_sep_thresholds	raw/swin_small/sep/sep_thresholds.csv	raw/swin_small/sep/sep_thresholds.csv	5	original
vit_base_patch14_dinov2_sep_sep_thresholds	raw/vit_base_patch14_dinov2/sep/sep_thresholds.csv	raw/vit_base_patch14_dinov2/sep/sep_thresholds.csv	5	original
vit_base_patch16_224_dino_sep_sep_thresholds	raw/vit_base_patch16_224_dino/sep/sep_thresholds.csv	raw/vit_base_patch16_224_dino/sep/sep_thresholds.csv	5	original
vit_base_patch16_224_mae_sep_sep_thresholds	raw/vit_base_patch16_224_mae/sep/sep_thresholds.csv	raw/vit_base_patch16_224_mae/sep/sep_thresholds.csv	5	original
vit_base_patch16_clip_openai_sep_sep_thresholds	raw/vit_base_patch16_clip_openai/sep/sep_thresholds.csv	raw/vit_base_patch16_clip_openai/sep/sep_thresholds.csv	5	original
vit_huge_patch14_224_mae_sep_sep_thresholds	raw/vit_huge_patch14_224_mae/sep/sep_thresholds.csv	raw/vit_huge_patch14_224_mae/sep/sep_thresholds.csv	5	original
vit_large_21k_in1k_sep_sep_thresholds	raw/vit_large_21k_in1k/sep/sep_thresholds.csv	raw/vit_large_21k_in1k/sep/sep_thresholds.csv	5	original
vit_large_patch14_clip_openai_sep_sep_thresholds	raw/vit_large_patch14_clip_openai/sep/sep_thresholds.csv	raw/vit_large_patch14_clip_openai/sep/sep_thresholds.csv	5	original
vit_large_patch14_dinov2_sep_sep_thresholds	raw/vit_large_patch14_dinov2/sep/sep_thresholds.csv	raw/vit_large_patch14_dinov2/sep/sep_thresholds.csv	5	original
vit_large_patch16_224_mae_sep_sep_thresholds	raw/vit_large_patch16_224_mae/sep/sep_thresholds.csv	raw/vit_large_patch16_224_mae/sep/sep_thresholds.csv	5	original
vit_small_patch16_224_dino_sep_sep_thresholds	raw/vit_small_patch16_224_dino/sep/sep_thresholds.csv	raw/vit_small_patch16_224_dino/sep/sep_thresholds.csv	5	original
vit_tiny_patch16_224_21k_sep_sep_thresholds	raw/vit_tiny_patch16_224_21k/sep/sep_thresholds.csv	raw/vit_tiny_patch16_224_21k/sep/sep_thresholds.csv	5	original

Source and intended use

The files are derived from the paper's representation-analysis workflow, including per-image SVD, dataset-level PCA, and Spectral Energy Pattern summaries. They are provided for inspecting the reported analyses and for regenerating tables or figures with the corresponding GitHub scripts. The artifacts are not a substitute for ImageNet-1K or for rerunning feature extraction.

License and third-party data

The repository content is released under apache-2.0. ImageNet images are not included; users remain responsible for the terms of ImageNet and all upstream software or model assets.

Citation

@inproceedings{tian2026encodingmismatch,
  title     = {From Per-Image Low-Rank to Encoding Mismatch:
               Rethinking Feature Distillation in Vision Transformers},
  author    = {Tian, Huiyuan and Xu, Bonan and Li, Shijian},
  booktitle = {Proceedings of the 43rd International Conference on Machine Learning},
  year      = {2026}
}