metadata
task_categories:
- image-text-to-text
language:
- en
dataset_info:
- config_name: attribute_grounding_and_alignment
features:
- name: idx
dtype: int32
- name: imgname
dtype: string
- name: chart_type
dtype: string
- name: attribute
dtype: string
- name: chart_1_grounding
dtype: string
- name: chart_2_grounding
dtype: string
- name: alignment_json
dtype: string
- name: image_pair
dtype: image
splits:
- name: test
num_bytes: 144273203.125
num_examples: 1751
download_size: 140665218
dataset_size: 144273203.125
- config_name: data_grounding_and_alignment
features:
- name: idx
dtype: int32
- name: imgname
dtype: string
- name: chart_type
dtype: string
- name: num_cell_difference
dtype: int32
- name: chart_1_grounding
dtype: string
- name: chart_2_grounding
dtype: string
- name: alignment_json
dtype: string
- name: image_pair
dtype: image
splits:
- name: test
num_bytes: 391677829.125
num_examples: 4287
download_size: 382856565
dataset_size: 391677829.125
- config_name: robustness
features:
- name: idx
dtype: int32
- name: set_idx
dtype: int32
- name: set_pair_idx
dtype: int32
- name: imgname
dtype: string
- name: chart_type
dtype: string
- name: num_cell_difference
dtype: int32
- name: attribute_varied
dtype: string
- name: alignment_json
dtype: string
- name: image_pair
dtype: image
splits:
- name: test
num_bytes: 1339741262.125
num_examples: 16575
download_size: 1277778734
dataset_size: 1339741262.125
configs:
- config_name: attribute_grounding_and_alignment
data_files:
- split: test
path: attribute_grounding_and_alignment/test-*
- config_name: data_grounding_and_alignment
data_files:
- split: test
path: data_grounding_and_alignment/test-*
- config_name: robustness
data_files:
- split: test
path: robustness/test-*
π ChartAlignBench
ChartAlignBench is a multi-modal benchmark designed to evaluate vision-language models (VLMs) on dense-level chart grounding and multi-chart alignment to comprehensively assess fine-grained chart understanding in VLMs.
π Overview
ChartAlignBench contains 9K+ instances, divided into three evaluation subsets:-
- Data Grounding & Alignment: Paired charts differ in underlying data values visualized by the chart.
- Attribute Grounding & Alignment: Paired charts differ in visualization attribute e.g., color, legend or text style.
- Robustness: Multiple variants of paired charts for Data Alignment, with different attributes across variants.
Examples of chart pair instances in ChartAlignBench.
π Instruction
For each subset, the test*.parquet files contain the annotations and image pairs pre-loaded for processing with HF Datasets.
from datasets import load_dataset
data_grounding_and_alignment_subset = load_dataset("umd-zhou-lab/ChartAlignBench", "data_grounding_and_alignment")
attribute_grounding_and_alignment_subset = load_dataset("umd-zhou-lab/ChartAlignBench", "attribute_grounding_and_alignment")
robustness_subset = load_dataset("umd-zhou-lab/ChartAlignBench", "robustness")
π Dataset Description
Common fields (all subsets)
| Field | Description |
|---|---|
| idx | Unique index of the sample in the set |
| imgname | Name of the chart image |
| chart_type | Type of chart |
| alignment_json | Alignment label for the chart pair |
| image_pair | Image of horizontally stitched chart pair (PIL.Image) |
Data Grounding and Alignment
| Field | Description |
|---|---|
| num_cell_difference | Number of data points which differ between the chart pair |
| chart_1_grounding | Grounding label for chart 1 (csv table) |
| chart_2_grounding | Grounding label for chart 2 (csv table) |
Attribute Grounding and Alignment
| Field | Description |
|---|---|
| attribute | Attribute which differs between the chart pair |
| chart_1_grounding | Grounding label for chart 1 (json) |
| chart_2_grounding | Grounding label for chart 2 (json) |
Robustness
| Field | Description |
|---|---|
| set_idx | Set index which groups the 5 chart pairs in a robustness set |
| set_pair_idx | Pair index within a set (1-5) |
| num_cell_difference | Number of data points which differ between the chart pair |
| attribute_varied | Attribute which varies across the 5 chart pairs |