llm-semantic-router/Vela-1.0-Encoder-307M-Embedding

Docs | Blog | Slack | GitHub

Vela Embedding

Vela Embedding turns text into vectors for semantic search, retrieval, and clustering.

307M parameters · Input capacity: 32,768 tokens, including special tokens.

The example returns a 768-dimensional unit vector for each text.

Evaluation

Compared with the original mmBERT Embedding on the same fixed development subsets. Scores are ×100; higher is better.

Evaluation	Metric	Original mmBERT	Vela
MIRACL · 320 queries	nDCG@10	74.51	75.16
Natural Questions · 128 queries	nDCG@10	96.36	95.66
SciFact · 48 queries	nDCG@10	64.78	56.36
QASPER · 64 queries	nDCG@10	33.54	28.60
Controlled 4K–32K context · 288 pairs	Pair accuracy	55.21	54.17
32K end-position subset · 24 pairs	Pair accuracy	50.00	45.83
STS-B · 1,500 pairs	Spearman	84.78	79.83
PAWS-X · 11,850 pairs	Accuracy¹	43.11	44.11

Controlled context · pair accuracy	Original mmBERT	Vela
4K · 72 pairs	52.78	52.78
8K · 72 pairs	56.94	56.94
16K · 72 pairs	58.33	55.56
32K · 72 pairs	52.78	51.39

Both models use 22 layers, 768 dimensions, FP32, identical tokenization and candidate pools, and complete inputs without truncation. MIRACL covers Arabic, Spanish, Japanese and Chinese (80 queries each). Controlled context places relevant text at different positions; the 32K end-position row is a subset. These development sets informed Vela checkpoint selection; they are not independent tests or full benchmark leaderboards. Original model training overlap is unknown.

¹ PAWS-X uses a fixed cosine threshold of 0.5.

Quick start

With SentenceTransformers and PyTorch:

from sentence_transformers import SentenceTransformer

model_id = "llm-semantic-router/Vela-1.0-Encoder-307M-Embedding"
model = SentenceTransformer(model_id, device="cpu")
vectors = model.encode(["The library opens in the morning.", "图书馆早上开门。"])
print(vectors.shape)  # (2, 768)

Explore the Vela model collection