alibayram/wikipedia-40-langs-with-embeddings-embeddinggemma2

Wikipedia 40 Languages with EmbeddingGemma-2 Retrieval Embeddings

Precomputed teacher embeddings from google/embeddinggemma-2 for offline embedding distillation, built on alibayram/wikipedia-40-langs. It is the distillation corpus for alibayram/embeddingmagibu2, following the pipeline of arXiv:2605.29992.

Status: splits are added as they finish (test → validation → train). Live progress and ETA: progress.json. Files under progress/ are partial chunks of the split being encoded and are removed once that split is pushed.

Splits

Split	Rows
test	58,000
validation	174,000
train	580,000

Columns

Column	Description
lang, title, url	From the source dataset
text	Article text, clipped so the document fits 8,192 tokens in both tokenizers (see below). This is exactly the text that was embedded.
truncated	true if text was clipped (~2.6% of rows)
original_text_chars	Character length of the unclipped article
embedgemma2_query_embedding	Teacher embedding of task: search result | query: {title} (768-d, L2-normalized)
embedgemma2_document_embedding	Teacher embedding of title: {title} | text: {text} (768-d, L2-normalized)

Both targets use EmbeddingGemma's retrieval prompt formats, so a student can learn query and document representations; the title query and its article form a natural positive pair.

Clipping to both tokenizers

The student model uses the Turkish-optimized magibu/Altus tokenizer, which needs ~21% fewer tokens for Turkish but up to ~15% more for some languages (e.g. Russian). To guarantee that student and teacher always see the same input, each rendered document (title: … | text: …) was clipped at the first character where either the teacher tokenizer or Altus exceeds 8,184 tokens (8,192 minus a reserve for special tokens), preferring a word boundary. No row exceeds 8,192 tokens in either tokenizer.

How it was made

• Teacher: google/embeddinggemma-2, bf16 on a single H200, mean pooling, normalized outputs. Verified against fp32 SentenceTransformer.encode (cosine ≥ 0.9997).
• Global length sorting with token-budget batching; script: prepare_distillation_dataset.py.

Usage

from datasets import load_dataset

ds = load_dataset("alibayram/wikipedia-40-langs-with-embeddings-embeddinggemma2", split="test")
row = ds[0]
doc_input = f"title: {row['title']} | text: {row['text']}"   # student input for the document target
query_input = f"task: search result | query: {row['title']}"  # student input for the query target

License

Text is from Wikipedia under CC BY-SA 4.0; embeddings are derived from it.