---
pretty_name: PAG Benchmark Data
task_categories:
- feature-extraction
tags:
- approximate-nearest-neighbor-search
- vector-search
- l2
- cosine
- mips
---

# PAG Benchmark Data

Benchmark data files for the Projection-Augmented Graph (PAG) project.

## Layout

```text
data/
  glove/
    base.fbin
    query.fbin
    gt1000.ibin
  sift/
    base.fbin
    query.fbin
    gt1000.ibin
  music/
    base.fbin
    query.fbin
    gt1000.ibin
```

## Datasets

| Directory | Metric | Base vectors | Queries | Dimension | Ground truth |
|---|---:|---:|---:|---:|---:|
| `data/glove` | L2 | 1,193,514 | 1,000 | 200 | top-1000 |
| `data/sift` | Cosine | 10,000,000 | 1,000 | 128 | top-1000 |
| `data/music` | MIPS | 1,000,000 | 10,000 | 100 | top-1000 |

## File Format

`base.fbin` and `query.fbin`:

```text
uint32 rows
uint32 dim
float32 data[rows][dim]
```

`gt1000.ibin`:

```text
uint32 rows
uint32 k
uint32 ids[rows][k]
```

## Usage

```bash
hf download ckadzh8/pag-benchmark-data --repo-type dataset --local-dir .
./scripts/run_glove.sh
./scripts/run_sift.sh
./scripts/run_music.sh
```

The ground-truth files store exact top-1000 neighbor IDs for the listed metric.