---
language:
- en
license: cc-by-nc-4.0
size_categories:
- 10K<n<100K
task_categories:
- image-to-text
- table-question-answering
- table-to-text
tags:
- synthetic
- chart
configs:
- config_name: encoder_training
  data_files:
  - split: train
    path: hf_ready/encoder_training/data/train-*.parquet
  - split: test
    path: hf_ready/encoder_training/data/test-*.parquet
- config_name: decoder_training
  data_files:
  - split: train
    path: hf_ready/decoder_training/data/train-*.parquet
  - split: test
    path: hf_ready/decoder_training/data/test-*.parquet
- config_name: pipeline_testing
  data_files:
  - split: test
    path: hf_ready/pipeline_testing/data/test-*.parquet
---

# Bar Chart Annotation Dataset

This dataset accompanies the paper [Bar-JEPA: Extracting Values from Bar Chart with Joint-Embedding Predictive Architecture](https://huggingface.co/papers/2608.06062). Code is available at [https://github.com/dralois/Bar-JEPA](https://github.com/dralois/Bar-JEPA).

A collection of synthetic bar chart images with detailed structural annotations. Split into three subsets for different stages of a chart understanding pipeline.

---

## Subsets

| Subset | Folder | Train | Test |
|---|---|---|---|
| `encoder_training` | `hf_ready/encoder_training/` | 100,000 | 100 |
| `decoder_training` | `hf_ready/decoder_training/` | 17,000 | 3,400 |
| `pipeline_testing` | `hf_ready/pipeline_testing/` | — | 100 |

Each subset is a standalone HuggingFace dataset (its own repo), formatted as sharded Parquet files under a `data/` subfolder.

---

## Columns

| Column | Type | Description |
|---|---|---|
| `filename` | `string` | Original file stem (e.g. `train_42`) — use to reconstruct or cross-reference source files |
| `image` | `binary` | Original PNG bytes (RGBA, 1272×954) |
| `annotation` | `string` | Verbatim JSON annotation — identical to the source `.json` files |
| `chart_type` | `string` | `"vertical"` or `"horizontal"` |
| `title` | `string` | Chart title text |
| `num_features` | `int32` | Number of data features (legend entries) |
| `num_categories` | `int32` | Number of category-axis ticks |
| `value_min` | `float32` | Value axis minimum |
| `value_max` | `float32` | Value axis maximum |

---

## Loading

```python
from datasets import load_dataset

ds = load_dataset("<org>/encoder_training")
sample = ds["train"][0]

# Image
from PIL import Image
from io import BytesIO
img = Image.open(BytesIO(sample["image"]))

# Annotation (full structure, unchanged)
import json
ann = json.loads(sample["annotation"])
# ann["chart_metadata"] — title, type, bar_width, legend, origin, size
# ann["data"]           — category_axis, value_axis, features
```

---

## Annotation Format

Each annotation follows this schema (see `format.json` for the full JSON Schema):

```
{
  "chart_metadata": {
    "title":     { "text": "...", "bbox": [x1, y1, x2, y2] },
    "type":      "vertical" | "horizontal",
    "bar_width": <int>,
    "legend": {
      "bbox": [...],
      "entries": [
        { "feature": { "text": "...", "bbox": [...] },
          "patch":   { "bbox": [...] } }
      ]
    },
    "origin": { "bbox": [...] },
    "size":   { "bbox": [...] }
  },
  "data": {
    "category_axis": {
      "label": { "text": "...", "bbox": [...] },
      "bbox":  [...],
      "ticks": [ { "label": { "text": "...", "bbox": [...] }, "bbox": [...] } ]
    },
    "value_axis": {
      "label":     { "text": "...", "bbox": [...] },
      "min_value": <number>,
      "max_value": <number>,
      "bbox":      [...],
      "ticks":     [ { "label": { "text": "...", "bbox": [...] }, "bbox": [...] } ]
    },
    "features": [
      {
        "feature": "...",
        "data": [
          { "category": "...", "value": <number>, "bbox": [...] }
        ]
      }
    ]
  }
}
```

All bounding boxes are `[x1, y1, x2, y2]` in pixel coordinates.