README.md

---
license: cc-by-nc-nd-4.0
task_categories:
  - image-to-text
  - document-question-answering
language:
  - en
size_categories:
  - n<1K
tags:
  - selection-detection
  - checkbox-detection
  - benchmark
  - document-ai
  - selection-f1
  - ocr
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*.parquet
  - config_name: results
    data_files:
      - split: train
        path: results/train-*.parquet
---

# PulseBench-Select

**A benchmark for selected-option detection in document images.**

PulseBench-Select contains 485 cleaned document images with ground-truth annotations for checkboxes, radio buttons, and marked answer choices. Each sample pairs a document image with public ground truth for the visible options that are selected.

- **Scoring methodology (GitHub):** `https://github.com/Pulse-Software-Corp/PulseBench-Select`

## Quick Start

```python
from datasets import load_dataset
import json

# Load benchmark data: document images plus cleaned ground truth.
ds = load_dataset("pulse-ai/PulseBench-Select")

sample = ds["train"][0]
sample["sample_id"]                    # Public sample id
sample["image"]                        # PIL image of the document page
gt = json.loads(sample["ground_truth"])
gt["selected_items"]                   # Selected options used for scoring

# Load aggregate benchmark results.
results = load_dataset("pulse-ai/PulseBench-Select", "results")

row = results["train"][0]
row["display_name"]                    # Provider display name
row["selection_f1"]                    # Corpus-pooled Selection F1
```

## Dataset Overview

| Split | Samples | Page Annotations | Selected Items | Selection Candidates |
|-------|---------|------------------|----------------|----------------------|
| train | 485 | 14,516 | 1,976 | 4,180 |

The benchmark focuses on pages where systems must determine which visible options are selected. Ground-truth coordinates are normalized eight-point polygons in reading order: `[x0, y0, x1, y1, x2, y2, x3, y3]`.

459 samples contain at least one selected item; 26 samples contain no selected items and are retained to measure false positives.

## Scoring: Selection F1

Selection F1 evaluates only the positive selected class.

1. **Parse** ground truth and predictions into selected-item records with `sample_id`, `page`, `content`, `bbox`, and `selected`.
2. **Match** each predicted selected item to the best unmatched ground-truth selected item on the same sample and page.
3. **Filter matches** with content token overlap >= 0.80. If both items include 8-point bounding boxes, the bbox centroid distance must also be <= 0.35 in normalized page units.
4. **Score** matched selected items as true positives, unmatched predictions as false positives, and unmatched ground-truth items as false negatives.
5. **Report** corpus-pooled micro precision, recall, and F1, along with per-sample macro diagnostics.

Token overlap is computed as:

```text
|tokens(ground_truth) intersect tokens(prediction)| / max(|tokens(ground_truth)|, |tokens(prediction)|)
```

The bbox centroid check is a veto used to prevent repeated labels with identical text, such as multiple `Yes` or `No` options on the same page, from matching the wrong spatial item. The public scorer supports disabling this check with `--centroid-max -1`.

## Results

We evaluated 6 systems using Selection F1. Scores below are corpus-pooled micro precision, recall, and F1 from the benchmark run associated with this release.

| Rank | Provider | Precision | Recall | Selection F1 |
|------|----------|-----------|--------|--------------|
| 1 | **Pulse** | **0.782** | **0.761** | **0.772** |
| 2 | GPT-5.5 | 0.383 | 0.311 | 0.343 |
| 3 | Gemini 3.1 Pro | 0.334 | 0.317 | 0.325 |
| 4 | Gemini 3.5 Flash | 0.317 | 0.311 | 0.314 |
| 5 | Claude Opus 4.8 | 0.307 | 0.293 | 0.300 |
| 6 | GPT-4o | 0.223 | 0.191 | 0.206 |

The `results` config includes these aggregate results plus macro precision, macro recall, and skipped-sample counts for each provider.

## Schema

### Default config

| Column | Type | Description |
|--------|------|-------------|
| `sample_id` | string | Stable public sample identifier |
| `image` | image | Document image |
| `ground_truth` | string | JSON with `page_count`, `annotations`, and `selected_items` |
| `annotation_count` | int | Number of cleaned page annotations |
| `selected_count` | int | Number of selected ground-truth items |
| `selection_candidate_count` | int | Number of annotations or cells containing visible selection marks |
| `selection_stats` | string | JSON summary for the row |

### Results config

| Column | Type | Description |
|--------|------|-------------|
| `rank` | int | Rank by corpus-pooled Selection F1 |
| `provider` | string | Provider identifier |
| `display_name` | string | Provider display name |
| `precision` | float | Corpus-pooled positive-class precision |
| `recall` | float | Corpus-pooled positive-class recall |
| `selection_f1` | float | Corpus-pooled positive-class F1 |
| `macro_precision` | float | Mean per-sample precision over scored samples |
| `macro_recall` | float | Mean per-sample recall over scored samples |
| `macro_skipped_samples` | int | Samples skipped from macro averaging because precision or recall was undefined |
| `metric_version` | string | Metric version used for the reported row |

## Ground Truth Format

```json
{
  "page_count": 1,
  "annotations": [
    {
      "category": "List-item",
      "bbox": [0.08, 0.33, 0.20, 0.33, 0.20, 0.35, 0.08, 0.35],
      "content": "B) Example option",
      "page": 1,
      "selected": true,
      "selection_candidate": true
    }
  ],
  "selected_items": [
    {
      "page": 1,
      "bbox": [0.08, 0.33, 0.20, 0.33, 0.20, 0.35, 0.08, 0.35],
      "content": "B) Example option",
      "category": "List-item"
    }
  ]
}
```

## License

This dataset is released under [CC BY-NC-ND 4.0](https://creativecommons.org/licenses/by-nc-nd/4.0/).