---
dataset_info:
  features:
  - name: image
    dtype: image
  - name: label
    dtype:
      class_label:
        names:
          '0': '0'
          '1': '1'
  splits:
  - name: train
    num_bytes: 83803600.0
    num_examples: 100000
  - name: validation
    num_bytes: 8609095.48
    num_examples: 10240
  - name: test
    num_bytes: 17218416.240000002
    num_examples: 20480
  download_size: 74270288
  dataset_size: 109631111.72
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: validation
    path: data/validation-*
  - split: test
    path: data/test-*
---

## Arrow Pointing Extrapolation

This dataset contains the exact images used for the extrapolation experiments in [pLSTM](https://huggingface.co/papers/2506.11997).
It is a synthetic dataset of arrows pointing to circles and should measure how well an image model can learn the classification 
'if the arrow points to the circle' at small (192x192) scales and extrapolate/generalize (without previous resizing of the image input) 
to larger scales (384x384).

Note that for the correct validation and test extrapolation subsets, you have to filter for the larger images:
```
from datasets import load_dataset
ds = load_dataset('ml-jku/arrow_pointing_extrapolation')
ds_val_ext = ds['validation'].filter(lambda sample: sample['image'].size == (384, 384))
ds_test_ext = ds['test'].filter(lambda sample: sample['image'].size == (384, 384))
```