---
dataset_info:
- config_name: eval
  features:
  - name: query
    dtype: string
  - name: image
    dtype: image
  - name: negative_0
    dtype: image
  - name: negative_1
    dtype: image
  - name: negative_2
    dtype: image
  - name: negative_3
    dtype: image
  splits:
  - name: train
    num_bytes: 356052075
    num_examples: 300
  download_size: 328567541
  dataset_size: 356052075
- config_name: full
  features:
  - name: query
    dtype: string
  - name: image
    dtype: image
  - name: negative_0
    dtype: image
  - name: negative_1
    dtype: image
  - name: negative_2
    dtype: image
  - name: negative_3
    dtype: image
  splits:
  - name: train
    num_bytes: 62370383388
    num_examples: 53512
  download_size: 57161629955
  dataset_size: 62370383388
- config_name: train
  features:
  - name: query
    dtype: string
  - name: image
    dtype: image
  - name: negative_0
    dtype: image
  - name: negative_1
    dtype: image
  - name: negative_2
    dtype: image
  - name: negative_3
    dtype: image
  splits:
  - name: train
    num_bytes: 11548548625
    num_examples: 10000
  download_size: 10570074118
  dataset_size: 11548548625
configs:
- config_name: eval
  data_files:
  - split: train
    path: eval/train-*
- config_name: full
  data_files:
  - split: train
    path: full/train-*
- config_name: train
  data_files:
  - split: train
    path: train/train-*
license: apache-2.0
language:
- en
pretty_name: Visual Document Retrieval Dataset
---

# llamaindex-vdr-en-train-preprocessed

This dataset is a preprocessed English subset of [llamaindex/vdr-multilingual-train](https://huggingface.co/datasets/llamaindex/vdr-multilingual-train), prepared for training multimodal [Sentence Transformer](https://sbert.net) embedding models on document screenshot retrieval.

## Changes from the original dataset

The original [llamaindex/vdr-multilingual-train](https://huggingface.co/datasets/llamaindex/vdr-multilingual-train) dataset stores hard negatives as a list of ID strings that reference other rows. This dataset makes two key changes:

1. **English only**: Only the English subset (53,512 samples) is included.
2. **Resolved negatives as images**: 4 out of the 16 hard negatives are resolved from IDs into the actual document screenshot images, stored as `negative_0` through `negative_3`. This makes the dataset directly usable for training with Sentence Transformers without any additional preprocessing.

## Dataset Structure

Each sample contains:

| Column | Type | Description |
|---|---|---|
| `query` | `string` | A synthetic text query associated with the document screenshot |
| `image` | `image` | The positive document screenshot (PDF page rendered as an image) |
| `negative_0` | `image` | Hard negative document screenshot (closest) |
| `negative_1` | `image` | Hard negative document screenshot |
| `negative_2` | `image` | Hard negative document screenshot |
| `negative_3` | `image` | Hard negative document screenshot (furthest) |

## Configs

| Config | Samples | Description |
|---|---|---|
| `full` | 53,512 | All English samples |
| `train` | 10,000 | The first 10,000 samples (0–9,999) from the `full` dataset |
| `eval` | 300 | The next 300 samples (10,000–10,299) from the `full` dataset |

You can certainly train on the `full` dataset, and then you're recommended to make your own eval/test splits. Do not combine `full` and `eval`, as you'll train and evaluate on the same data.

## Usage

```python
from datasets import load_dataset

# Load the training split
train_dataset = load_dataset("tomaarsen/llamaindex-vdr-en-train-preprocessed", "train", split="train")

# Load the evaluation split
eval_dataset = load_dataset("tomaarsen/llamaindex-vdr-en-train-preprocessed", "eval", split="train")

# Load all English samples
full_dataset = load_dataset("tomaarsen/llamaindex-vdr-en-train-preprocessed", "full", split="train")
```

## Training

This dataset can be used to finetune a multimodal Sentence Transformer model for document screenshot embedding. See the [training example](https://github.com/huggingface/sentence-transformers/blob/main/examples/sentence_transformer/training/multimodal/training_document_screenshot_embedding.py) for a full training script.

## Source

This dataset is derived from [llamaindex/vdr-multilingual-train](https://huggingface.co/datasets/llamaindex/vdr-multilingual-train), which consists of 500k multilingual query-image samples collected and generated from public internet PDFs. Queries were synthetically generated using VLMs (gemini-1.5-pro and Qwen2-VL-72B). See the original dataset card for full details on the data collection and curation process.