SIFT1B - Sharded DiskANN Indices

Pre-built DiskANN indices for the SIFT1B (BigANN) dataset, sharded for distributed vector search.

Dataset Info

• Source: BigANN Benchmarks
• Vectors: 1,000,000,000 (1 billion)
• Dimensions: 128
• Data type: uint8
• Queries: 10,000
• Distance: L2

DiskANN Parameters

• R (graph degree): 64
• L (build beam width): 100
• PQ bytes: 32

Shard Configurations

• shard_2: 2 shards x 500,000,000 vectors
• shard_3: 3 shards x ~333,333,333 vectors
• shard_5: 5 shards x 200,000,000 vectors
• shard_7: 7 shards x ~142,857,142 vectors
• shard_10: 10 shards x 100,000,000 vectors

File Structure

fbin/
  base.u8bin                          # Base vectors (1B x 128 uint8)
  queries.u8bin                       # Query vectors (10K x 128 uint8)
diskann/
  gt_100.bin                          # Ground truth (100-NN)
  shard_N/                            # N-shard configuration
    sift1b_64_100_32.shardX_disk.index                    # DiskANN disk index
    sift1b_64_100_32.shardX_disk.index_512_none.indices   # MinIO graph indices
    sift1b_64_100_32.shardX_disk.index_base_none.vectors  # MinIO vector data
    sift1b_base.shardX.fbin                               # Shard base data

Chunked Files

Files larger than 49 GB are split into chunks for upload:

• *_512_none.indices.part00, .part01, etc.
• *_base_none.vectors.part00, etc. (if applicable)
• fbin/base.u8bin.part00, etc.

To reassemble: cat file.part00 file.part01 ... > file

Usage

Download with huggingface_hub

from huggingface_hub import hf_hub_download

# Download a specific shard file
index = hf_hub_download(
    repo_id="maknee/sift1b",
    filename="diskann/shard_10/sift1b_64_100_32.shard0_disk.index",
    repo_type="dataset"
)

Download with git-lfs

git lfs install
git clone https://huggingface.co/datasets/maknee/sift1b

License

Same as source dataset (BigANN Benchmarks).