Neurosense

Neurosense is an embedding model optimized for semantic retrieval.

Training Summary

• Base model: sentence-transformers/all-MiniLM-L6-v2
• Training file: data/sample/train.jsonl
• Pair examples: 2
• Triplet examples: 2
• Epochs: 1
• Batch size: 2
• Max sequence length: 512

Usage

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("path-or-repo-id")
embeddings = model.encode([
    "example query",
    "example document"
], normalize_embeddings=True)

Intended Use

• Semantic search
• Dense retrieval / RAG
• Similarity matching

Retrieval Benchmark (No Generation)

Evaluated on February 27, 2026 in an isolated .test workflow using internet corpora from:

• mteb/fiqa
• mteb/nfcorpus
• mteb/scifact

Benchmark protocol:

• 150 held-out semantic search queries total (50 per corpus)
• Retrieval-only metrics (hit@1, hit@5, mrr@10, recall@10, map@10)
• No generation used at any step

Best checkpoint in the tuning run: baseline Neurosense checkpoint (models/Neurosense).

Aggregate metrics on the 150-query benchmark:

• hit@1: 0.6667
• hit@5: 0.7933
• mrr@10: 0.7242
• recall@10: 0.6249
• map@10: 0.5691

Detailed local benchmark artifacts were written under .test/results/.