Update README.md

README.md

@@ -2,4 +2,128 @@
 license: other
 license_name: link-attribution
 license_link: https://dejan.ai/link-attribution/
+pretty_name: 200k Wikidata Embeddings (Gemini)
+language:
+- en
+tags:
+- wikidata
+- embeddings
+- semantic-search
+- entity-linking
+- gemini
+task_categories:
+- feature-extraction
+- sentence-similarity
+size_categories:
+- 100K<n<1M
+# license: choose one before publishing (see "Licensing" below)
+configs:
+- config_name: default
+  data_files:
+  - split: train
+    path: gemini_norm.parquet
+dataset_info:
+  features:
+  - name: entity_id
+    dtype: string
+  - name: label
+    dtype: string
+  - name: embedding
+    sequence: float32
+    length: 768
+  splits:
+  - name: train
+    num_examples: 199998
+  config_name: default
 ---
+
+# 200k Wikidata Embeddings (Gemini)
+
+Text embeddings for ~200,000 Wikidata entities, generated from each entity's
+English label with Google's **`gemini-embedding-001`** model. Vectors are
+768-dimensional and **L2-normalized to unit length**, so cosine similarity
+equals a plain dot product.
+
+## What's inside
+
+| Column | Type | Description |
+|---|---|---|
+| `entity_id` | `string` | Wikidata Q-identifier, e.g. `Q42` |
+| `label` | `string` | English label, e.g. `Douglas Adams` |
+| `embedding` | `list<float32>[768]` | Unit-length embedding of the label |
+
+- **Rows:** 199,998
+- **Dimensions:** 768
+- **Normalization:** L2 (every vector has norm 1.0)
+- **Source model:** `gemini-embedding-001` (native 3072-dim, requested at
+  `output_dimensionality=768`, then L2-normalized)
+- **File:** `gemini_norm.parquet`
+
+The entities are a sample of the full Wikidata label set (~118.7M labelled
+entities). Only the label string was embedded — no descriptions, aliases, or
+statements were used.
+
+## Usage
+
+```python
+from datasets import load_dataset
+
+ds = load_dataset("dejanseo/200k-wiki-data-embeddings-gemini", split="train")
+print(ds[0]["entity_id"], ds[0]["label"])
+print(len(ds[0]["embedding"]))  # 768
+```
+
+Semantic search with cosine similarity (a dot product, since vectors are
+unit-length):
+
+```python
+import numpy as np
+
+emb = np.array(ds["embedding"], dtype=np.float32)   # (199998, 768)
+ids = ds["entity_id"]
+labels = ds["label"]
+
+def nearest(query_vec, k=10):
+    scores = emb @ np.asarray(query_vec, dtype=np.float32)  # cosine similarity
+    top = np.argsort(-scores)[:k]
+    return [(ids[i], labels[i], float(scores[i])) for i in top]
+```
+
+## How it was built
+
+1. English labels were pulled from a local Wikidata label database.
+2. Each label was embedded via the Gemini Batch Embeddings API
+   (`gemini-embedding-001`, `output_dimensionality=768`).
+3. The returned vectors were **L2-normalized** to unit length and written to
+   Parquet.
+
+Because the vectors are truncated Matryoshka outputs (768 of the model's native
+3072 dims) and then renormalized, they are directly usable for cosine /
+dot-product retrieval.
+
+## Companion dataset
+
+A parallel set of embeddings for the **same 199,998 entities** was produced with
+the open `google/embeddinggemma-300m` model (also 768-dim, unit-length). Because
+the `entity_id` keys are identical, the two sets can be joined row-for-row for
+cross-model comparison.
+
+> Note on cross-model use: the two models place entities in differently-oriented
+> spaces. Do **not** compare a Gemini vector directly against a Gemma vector —
+> raw cross-model similarity is meaningless. Within a single model, similarity is
+> well-behaved.
+
+## Licensing
+
+Wikidata labels are released under **CC0**. The embedding vectors are derived
+outputs of Google's Gemini API and are subject to Google's applicable terms.
+Choose and set a `license:` value in the metadata above that reflects how you
+intend to distribute the derived vectors before publishing.
+
+## Citation
+
+If you use this dataset, please cite the source model and Wikidata:
+
+- Google, `gemini-embedding-001`
+- Wikidata (Wikimedia Foundation)
+