metadata
license: cc0-1.0
pretty_name: The Met Open Access — siglip2 embeddings (OEmbeddings spec)
tags:
- art
- museum
- embeddings
- oembeddings
- siglip2
configs:
- config_name: default
data_files:
- split: train
path: default/train/siglip2-oembeddings-*.parquet
metmuseum/openaccess-embeddings-siglip2-oembeddings
Embeddings produced by google/siglip2-so400m-patch14-384 over metmuseum/openaccess, shaped to match the
go-embeddingsdb Record
struct (with the OEmbeddings spec
inside attributes).
| Column | Type | Notes |
|---|---|---|
provider |
string | "metmuseum" |
subject_id |
string | metmuseum:{objectID} — the artwork |
depiction_id |
string | metmuseum:{objectID}:primaryImage — the image |
model |
string | HF model id |
embeddings |
list<float32> | L2-normalised, dim = 1152 |
created |
int64 | Unix epoch seconds when this batch was generated |
attributes |
map<string, string> | OEmbeddings spec fields (see below) |
attributes struct
| Field | Required | Source |
|---|---|---|
type |
yes | always "image" |
preview |
yes (URI) | primaryImageSmall or primaryImage |
subject_url |
yes | objectURL (the metmuseum.org page) |
subject_title |
yes | Title |
subject_creditline |
yes | creditLine |
provider_name |
yes | "The Metropolitan Museum of Art" |
provider_url |
yes | https://www.metmuseum.org/ |
depiction_url |
optional | primaryImage |
Loading
from datasets import load_dataset
ds = load_dataset("metmuseum/openaccess-embeddings-siglip2-oembeddings", split="train")
row = ds[0]
print(row["subject_id"], row["model"], row["attributes"]["subject_title"])