---
language:
- en
license: apache-2.0
task_categories:
- image-text-to-text
pretty_name: ITT-Purpose
size_categories:
- 100K<n<1M
tags:
- multimodal
- image-text-to-text
- ocr
- table-qa
- latex
- vlm
- benchmark
dataset_info:
  features:
  - name: id
    dtype: string
  - name: image
    dtype: image
  - name: prompt
    dtype: string
  - name: response
    dtype: string
  - name: style
    dtype: string
  splits:
  - name: train
    num_examples: 100
  config_name: default
---

# ITT-Purpose

**Author:** convence

**ITT-Purpose** is a premium, hard, and clean benchmark dataset of **100** unique samples 
for training and evaluating image-to-text-to-text (Vision-Language) models.

## Dataset Structure

Each sample contains:
- `id`: A unique UUID string identifying the sample.
- `image`: The rendered visual document containing styled text, code configs, or structured tables.
- `prompt`: A high-difficulty instruction requesting visual layout parsing, math calculating, or semantic reasoning.
- `response`: The clean, correct ground truth text.
- `style`: One of three styles (`meaning`, `formatting`, `table`).

## Styles Covered

1. **Meaning**: Renders complex technical document segments with multi-hop semantic reasoning questions.
2. **Text Formatting**: Renders nested JSON, YAML configs, and Python functions, demanding code structure and detail extraction.
3. **Table**: Renders dense telemetry data tables with borders, demanding cell lookups, calculated aggregates, or full markdown table generation.

## Usage

```python
from datasets import load_dataset

ds = load_dataset("convence/ITT-Purpose", split="train")
print(ds[0])
```

## License

Apache 2.0