olanokhin/hnsw-turboquant-glove

TurboQuant x HNSW: GloVe scaling curve

This Space benchmarks a TurboQuant-style vector compression pipeline on real GloVe-200 embeddings:

• —randomized Hadamard rotation
• —Lloyd-Max scalar quantization
• —HNSW cosine search
• —Recall@10 vs compression across 1, 2, 4, 6, and 8 bits

Default benchmark setup:

• —dataset: olanokhin/glove-6b-200d-vectors
• —native dimension: 200
• —encoded dimension: 256 after Hadamard padding
• —corpus vectors: 25,000
• —queries: 500
• —HNSW: ef_construction=200, M=16

The point is not to claim a production implementation. The app reconstructs float32 vectors from compressed codes and builds a standard HNSW index. A production vector database would store codes directly and score through codebook lookup with SIMD kernels.

Dataset

For Hugging Face Spaces, set these environment variables:

HF_DATASET_ID=olanokhin/glove-6b-200d-vectors
HF_DATASET_SPLIT=train
HF_VECTOR_COLUMN=vector

Expected dataset schema:

• —split: train
• —vector column: list of 200 floats
• —optional word column: ignored by the benchmark

For a local real-GloVe run before uploading the dataset, download and unzip Stanford glove.6B.200d.txt, then launch with:

GLOVE_TXT_PATH=/path/to/glove.6B.200d.txt python app.py

Resume bullet

Benchmarked TurboQuant (Google ICLR 2026) on GloVe-200: mapped Recall@10 vs compression across 1-8 bit quantization; identified 8-bit as safe sweet spot (3.12x smaller, -0.7pp Recall@10) and 6-bit as aggressive sweet spot (4.17x smaller, -1.9pp).