YoKONCy/Cohere-1M-wikipedia-768d

Cohere-1M Wikipedia 768-d Embeddings

Pre-computed 768-dimensional embeddings of 1M English Wikipedia articles, generated using Cohere's embed-english-v2.0 (multilingual-22-12) model.

This dataset is used for ANN (Approximate Nearest Neighbor) benchmarking in the QuIVer paper (PVLDB Vol. 20, 2027).

Files

File	Shape	Format	Description
cohere_train.f32	1,000,000 × 768	float32 raw binary	Base vectors (L2-normalized)
cohere_test.f32	1,000 × 768	float32 raw binary	Query vectors (L2-normalized)
cohere_groundtruth.i32	1,000 × 1,000	int32 raw binary	Ground truth top-1000 neighbor IDs (cosine)

Usage

import numpy as np
from huggingface_hub import hf_hub_download

# Download files
train_path = hf_hub_download("YoKONCy/Cohere-1M-wikipedia-768d", "cohere_train.f32")
test_path = hf_hub_download("YoKONCy/Cohere-1M-wikipedia-768d", "cohere_test.f32")
gt_path = hf_hub_download("YoKONCy/Cohere-1M-wikipedia-768d", "cohere_groundtruth.i32")

# Load
train = np.fromfile(train_path, dtype=np.float32).reshape(-1, 768)
test = np.fromfile(test_path, dtype=np.float32).reshape(-1, 768)
gt = np.fromfile(gt_path, dtype=np.int32).reshape(1000, -1)

print(f"Train: {train.shape}, Test: {test.shape}, GT: {gt.shape}")

Data Format

All files use headerless raw binary format:

• .f32: contiguous float32 values, row-major. File size = N × D × 4 bytes.
• .i32: contiguous int32 values, row-major. File size = Q × K × 4 bytes.

Source

Originally sampled from Cohere/wikipedia-22-12-en-embeddings (now deprecated on HuggingFace). Vectors are L2-normalized for cosine similarity evaluation.

Citation

@article{quiver2026,
  title   = {QuIVer: Rethinking ANN Graph Topology via Training-Free Binary Quantization},
  author  = {Xiao, Wenxuan and Wang, Zhiyou and Li, Chengcheng},
  journal = {arXiv preprint arXiv:2605.02171},
  year    = {2026},
  url     = {https://arxiv.org/abs/2605.02171}
}