multimodal_embedding
multimodal-embedding-100M
Multimodal Embedding 100M
This dataset contains a 100M-row multimodal embedding corpus generated from LAION-style image-text data exported with img2dataset as WebDataset shards. Images were resized to 256 during the WebDataset creation step before embedding generation. The dataset is intended for large-scale vector database ingestion, ANN index construction, nearest-neighbor search, and retrieval benchmark experiments.
The dataset is stored as Parquet files and organized to keep… See the full description on the dataset page: https://huggingface.co/datasets/VDBBench/multimodal-embedding-100M.openonerec_multimodal_embedding
Original Data
Basic Information
Table schema:
pid BIGINT
vision_emb ARRAY<ARRAY<DOUBLE>>
text_emb ARRAY<DOUBLE>
Notes:
vision_emb is the image embedding
text_emb is the text embedding
Coverage Summary
The table currently contains 17,433,569 pids in total.
Breakdown:
Pids with both image and text embeddings: 15,647,227
Pids with image embedding only: 1,411,004
Pids with text embedding only: 375,338
In short:
Most pids already have complete multimodal… See the full description on the dataset page: https://huggingface.co/datasets/OpenOneRec/openonerec_multimodal_embedding.multimodal-embedding-10M
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… See the full description on the dataset page: https://huggingface.co/datasets/VDBBench/multimodal-embedding-10M.multimodal-embedding-1M
Multimodal Embedding 1M Benchmark Dataset
A vector search benchmark dataset containing 1M base vectors and 10K query vectors with pre-computed ground truth, generated from multimodal (image + text) inputs.
Dataset Description
Each embedding is produced by encoding an image-text pair into a single 4096-dimensional vector using Qwen3-VL-Embedding-8B, a state-of-the-art multimodal embedding model.
Source data: pixparse/cc3m-wds (Conceptual Captions 3M in WebDataset… See the full description on the dataset page: https://huggingface.co/datasets/VDBBench/multimodal-embedding-1M.t5-gemma-2-multimodal-embeddingslay-the-spire-2-card-multimodal-embeddings
Slay the Spire 2: Multimodal Card Embeddings
Joint text+image embeddings for every card in Slay the Spire 2 (Early Access), produced by Qwen/Qwen3-VL-Embedding-2B. One unit-normalized 1024-D vector per card. Mechanically AND visually similar cards land near each other; cards across STS1 and STS2 share the coordinate system.
This is the multimodal-embeddings dataset. For text-only embeddings or the underlying card metadata + portraits, see:
t22000t/slay-the-spire-2-cards - metadata… See the full description on the dataset page: https://huggingface.co/datasets/t22000t/slay-the-spire-2-card-multimodal-embeddings.