---
task_categories:
  - feature-extraction
language:
  - en
license: apache-2.0
size_categories:
  - 10M<n<100M
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*.parquet
      - split: test
        path: data/test-*.parquet
      - split: neighbors
        path: data/neighbors.parquet
dataset_info:
  features:
    - name: id
      dtype: int64
    - name: emb
      sequence: float32
  splits:
    - name: train
      num_examples: 10000000
    - name: test
      num_examples: 10000
    - name: neighbors
      num_examples: 10000
tags:
  - multimodal
  - embeddings
  - vector-search
  - benchmark
  - image-text
---

# Multimodal Embedding 10M Benchmark Dataset

A large-scale vector search benchmark dataset containing **10M base vectors** and **10K query vectors** with pre-computed ground truth (top-100 nearest neighbors by Inner Product), generated from multimodal (image + text) inputs.

## Dataset Summary

| Property | Value |
|----------|-------|
| Base Vectors | 10,000,000 |
| Query Vectors | 10,000 |
| Dimension | 4,096 |
| Distance Metric | Inner Product (IP) |
| Top-K Ground Truth | 100 |
| Vector dtype | float32 |
| Embedding Model | [Qwen3-VL-Embedding-8B](https://huggingface.co/Qwen/Qwen3-VL-Embedding-8B) |
| Source Data | [pixparse/cc12m-wds](https://huggingface.co/datasets/pixparse/cc12m-wds) (Conceptual Captions 12M) |
| Base/Query Overlap | Yes (0.1%, query shards 40-41 included in base) |

## Splits

| Split | Rows | Description |
|-------|------|-------------|
| `train` | 10,000,000 | Base vectors (20 parquet shards, 500K rows each) |
| `test` | 10,000 | Query vectors |
| `neighbors` | 10,000 | Ground truth: top-100 nearest neighbor IDs per query |

## Schema

**train / test:**
| Column | Type | Description |
|--------|------|-------------|
| `id` | int64 | Sequential identifier (0-indexed) |
| `emb` | list\<float32\> | 4096-dim L2-normalized embedding |

**neighbors:**
| Column | Type | Description |
|--------|------|-------------|
| `id` | int64 | Query ID (matches test split) |
| `neighbors` | list\<int64\> | Top-100 base vector IDs by IP score (descending) |

## Usage

```python
from datasets import load_dataset
import numpy as np

ds = load_dataset("WenxingZhu/multimodal-embedding-10M")

# Base and query embeddings
base_emb = np.array(ds["train"]["emb"])    # (10_000_000, 4096) - streams from 20 shards
query_emb = np.array(ds["test"]["emb"])    # (10_000, 4096)

# Ground truth neighbors
neighbors = ds["neighbors"]["neighbors"]   # list of 10K lists, each 100 int64 IDs
```

### Streaming (recommended for 10M)

```python
from datasets import load_dataset

ds = load_dataset("WenxingZhu/multimodal-embedding-10M", split="train", streaming=True)
for row in ds:
    emb = row["emb"]  # list of 4096 floats
    break
```

## Generation Details

- **Model**: Qwen3-VL-Embedding-8B via vLLM (pooling runner, bfloat16)
- **Input**: Image-text pairs from cc12m-wds WebDataset shards, streamed from HuggingFace
- **Normalization**: L2 normalized post-embedding
- **Hardware**: 6x NVIDIA A100-80GB (DGX)
- **Throughput**: ~25 samples/sec per GPU
- **Total Generation Time**: ~3 days for 10M embeddings
- **Ground Truth**: Brute-force inner product, chunked numpy computation

## Related

- [multimodal-embedding-1M](https://huggingface.co/datasets/WenxingZhu/multimodal-embedding-1M) — 1M version from cc3m-wds

## License

Apache 2.0