---
task_categories:
- text-retrieval
- image-to-text
- sentence-similarity
language:
- en
tags:
- embeddings
- vector-database
- benchmark
---

# GAS Indexing Artifacts

## Dataset Description

This dataset contains pre-computed deterministic centroids and associated geometric metadata generated using our GAS (Geometry-Aware Selection) algorithm. 
These artifacts are designed to benchmark Approximate Nearest Neighbor (ANN) search performance in privacy-preserving or dynamic vector database environments.

### Purpose

To serve as a standardized benchmark resource for evaluating the efficiency and recall of vector databases implementing the GAS architecture. 
It is specifically designed for integration with VectorDBBench.

### Dataset Summary

- **Source Data**: 
  - Wikipedia (Public Dataset)
  - LAION0400M (Public Dataset)
- **Embedding Model**: 
  - google/embeddinggemma-300m
  - sentence-transformers/clip-ViT-B-32

## Dataset Structure

For each embedding model, the directory contains two key file:

| Data | Description |
|-------|-------------|
| `centroids.npy` | centroids as followed IVF |

## Data Fields

### Centroids: `centroids.npy`

- **Purpose**: Finding the nearest clusters for IVF (Inverted File Index)
- **Type**: NumPy array (`np.ndarray`)
- **Shape**: `[32768, 768]` or `[1024, 512]`
- **Description**: 768-dimensional vectors representing 32,768 cluster centroids, or 512-dimensional vectors representing 1,024 cluster centroids.
- **Normalization**: L2-normalized (unit norm)
- **Format**: float32


## Dataset Creation

### Source Data

Source dataset is a large public dataset: 
- Wikipedia: [mixedbread-ai/wikipedia-data-en-2023-11](https://huggingface.co/datasets/mixedbread-ai/wikipedia-data-en-2023-11)
- LAION: [LAION-400M](https://laion.ai/blog/laion-400-open-dataset/).

### Preprocessing

1. Create Centroids by GAS approach:

   Description TBD

2. Chunking (for text): For texts exceeding 2048 tokens:

    - Split into chunks with ~100 token overlap
    - Embedded each chunk separately
    - Averaged chunk embeddings for final representation
   
3. Normalization: All embeddings are L2-normalized

### Embedding Generation

- Text:
  - Model: google/embeddinggemma-300m
  - Dimension: 768
  - Max Token Length: 2048
  - Normalization: L2-normalized

- Multi-Modal:
  - Model: sentence-transformers/clip-ViT-B-32
  - Dimension: 512
  - Normalization: L2-normalized

## Usage

```python
import wget

def download_centroids(embedding_model: str, dataset_dir: str) -> None:
    """Download pre-computed centroids for IVF_GAS."""
    dataset_link = f"https://huggingface.co/datasets/cryptolab-playground/gas-centroids/resolve/main/{embedding_model}"
    wget.download(f"{dataset_link}/centroids.npy", out="centroids.npy")
```

## License

Apache 2.0

## Citation

If you use this dataset, please cite:

```bibtex
@dataset{gas-centroids,
  author = {CryptoLab, Inc.},
  title = {GAS Centroids},
  year = {2025},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/cryptolab-playground/gas-centroids}
}
```

### Source Dataset Citation

```bibtex
@dataset{wikipedia_data_en_2023_11,
  author = {mixedbread-ai},
  title = {Wikipedia Data EN 2023 11},
  year = {2023},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/mixedbread-ai/wikipedia-data-en-2023-11}
}
```

```bibtex
@dataset{laion400m,
  author = {Schuhmann, Christoph and others},
  title = {LAION-AI},
  year = {2021},
  publisher = {LAION},
  url = {https://laion.ai/blog/laion-400-open-dataset}
}
```

### Embedding Model Citation

```bibtex
@misc{embeddinggemma,
  title={Embedding Gemma},
  author={Google},
  year={2024},
  url={https://huggingface.co/google/embeddinggemma-300m}
}
```

```bibtex
@misc{clipvitb32,
  title={CLIP ViT-B/32},
  author={Open AI},
  year={2021},
  url={https://huggingface.co/sentence-transformers/clip-ViT-B-32}
}
```

### Acknowledgments

- Original dataset: 
  - mixedbread-ai/wikipedia-data-en-2023-11
  - LAION-400M
- Embedding model: 
  - google/embeddinggemma-300m
  - sentence-transformers/clip-ViT-B-32
- Benchmark framework: VectorDBBench