---
language:
- en
license: mit
size_categories:
- 1K<n<10K
task_categories:
- image-classification
pretty_name: Metashift subset for PCBM reproduction
viewer: false
dataset_info:
- config_name: cherrypicked_task_1_bed_cat_dog
  features:
  - name: image
    dtype: image
  - name: label
    dtype:
      class_label:
        names:
          '0': airplane
          '1': bed
          '2': car
          '3': cow
          '4': keyboard
  splits:
  - name: train
    num_bytes: 28494
    num_examples: 500
  - name: test
    num_bytes: 28486
    num_examples: 500
  download_size: 477673284
  dataset_size: 56980
- config_name: cherrypicked_task_1_bed_dog_cat
  features:
  - name: image
    dtype: image
  - name: label
    dtype:
      class_label:
        names:
          '0': airplane
          '1': bed
          '2': car
          '3': cow
          '4': keyboard
  splits:
  - name: train
    num_bytes: 28490
    num_examples: 500
  - name: test
    num_bytes: 28478
    num_examples: 500
  download_size: 477673272
  dataset_size: 56968
- config_name: cherrypicked_task_2_table_books_cat
  features:
  - name: image
    dtype: image
  - name: label
    dtype:
      class_label:
        names:
          '0': beach
          '1': computer
          '2': motorcycle
          '3': stove
          '4': table
  splits:
  - name: train
    num_bytes: 28413
    num_examples: 500
  - name: test
    num_bytes: 28478
    num_examples: 500
  download_size: 477673223
  dataset_size: 56891
- config_name: cherrypicked_task_2_table_books_dog
  features:
  - name: image
    dtype: image
  - name: label
    dtype:
      class_label:
        names:
          '0': beach
          '1': computer
          '2': motorcycle
          '3': stove
          '4': table
  splits:
  - name: train
    num_bytes: 28411
    num_examples: 500
  - name: test
    num_bytes: 28477
    num_examples: 500
  download_size: 477673220
  dataset_size: 56888
- config_name: cherrypicked_task_2_table_cat_dog
  features:
  - name: image
    dtype: image
  - name: label
    dtype:
      class_label:
        names:
          '0': beach
          '1': computer
          '2': motorcycle
          '3': stove
          '4': table
  splits:
  - name: train
    num_bytes: 28477
    num_examples: 500
  - name: test
    num_bytes: 28485
    num_examples: 500
  download_size: 477673292
  dataset_size: 56962
- config_name: cherrypicked_task_2_table_dog_cat
  features:
  - name: image
    dtype: image
  - name: label
    dtype:
      class_label:
        names:
          '0': beach
          '1': computer
          '2': motorcycle
          '3': stove
          '4': table
  splits:
  - name: train
    num_bytes: 28476
    num_examples: 500
  - name: test
    num_bytes: 28484
    num_examples: 500
  download_size: 477673290
  dataset_size: 56960
- config_name: seed42_task_1_bed_cat_dog
  features:
  - name: image
    dtype: image
  - name: label
    dtype:
      class_label:
        names:
          '0': airplane
          '1': bed
          '2': car
          '3': cow
          '4': keyboard
  splits:
  - name: train
    num_bytes: 28498
    num_examples: 500
  - name: test
    num_bytes: 28480
    num_examples: 500
  download_size: 477673282
  dataset_size: 56978
- config_name: seed42_task_1_bed_dog_cat
  features:
  - name: image
    dtype: image
  - name: label
    dtype:
      class_label:
        names:
          '0': airplane
          '1': bed
          '2': car
          '3': cow
          '4': keyboard
  splits:
  - name: train
    num_bytes: 28501
    num_examples: 500
  - name: test
    num_bytes: 28485
    num_examples: 500
  download_size: 477673290
  dataset_size: 56986
- config_name: seed42_task_2_table_books_cat
  features:
  - name: image
    dtype: image
  - name: label
    dtype:
      class_label:
        names:
          '0': beach
          '1': computer
          '2': motorcycle
          '3': stove
          '4': table
  splits:
  - name: train
    num_bytes: 28434
    num_examples: 500
  - name: test
    num_bytes: 28481
    num_examples: 500
  download_size: 477673247
  dataset_size: 56915
- config_name: seed42_task_2_table_books_dog
  features:
  - name: image
    dtype: image
  - name: label
    dtype:
      class_label:
        names:
          '0': beach
          '1': computer
          '2': motorcycle
          '3': stove
          '4': table
  splits:
  - name: train
    num_bytes: 28434
    num_examples: 500
  - name: test
    num_bytes: 28479
    num_examples: 500
  download_size: 477673245
  dataset_size: 56913
- config_name: seed42_task_2_table_cat_dog
  features:
  - name: image
    dtype: image
  - name: label
    dtype:
      class_label:
        names:
          '0': beach
          '1': computer
          '2': motorcycle
          '3': stove
          '4': table
  splits:
  - name: train
    num_bytes: 28465
    num_examples: 500
  - name: test
    num_bytes: 28479
    num_examples: 500
  download_size: 477673274
  dataset_size: 56944
- config_name: seed42_task_2_table_dog_cat
  features:
  - name: image
    dtype: image
  - name: label
    dtype:
      class_label:
        names:
          '0': beach
          '1': computer
          '2': motorcycle
          '3': stove
          '4': table
  splits:
  - name: train
    num_bytes: 28463
    num_examples: 500
  - name: test
    num_bytes: 28481
    num_examples: 500
  download_size: 477673274
  dataset_size: 56944
---

# PCBM Metashift

For the sake of reproducibility, this dataset hosts the postprocessed Metashift according to [[Yuksekgonul et al.]](https://arxiv.org/pdf/2205.15480.pdf) for the use in Post-Hoc Concept Bottleneck Models. 


| Config Name | Description |
|---|---|
| `task_1_bed_cat_dog` | Task 1: bed(cat) -> bed(dog)  |
| `task_1_bed_dog_cat` | Task 1: bed(dog) -> bed(cat)  |
| `task_2_table_books_cat` | Task 2: table(books) -> table(cat)  |
| `task_2_table_books_dog` | Task 2: table(books) -> table(dog) |
| `task_2_table_cat_dog` | Task 2: table(cat) -> table(dog) |
| `task_2_table_dog_cat` | Task 2: table(dog) -> table(cat) | 

The script to generate this dataset can be found at  `scripts/generate.py`. You will need to download the [Metashift repo](https://github.com/Weixin-Liang/MetaShift) and the [Visual Genome dataset](https://nlp.stanford.edu/data/gqa/images.zip) as instructed in the Metashift repo.