Neurosense Compact GTE R1

Compact embedding model for semantic retrieval, designed to stay under 200MB while improving retrieval quality over the previous Neurosense checkpoint.

Model Summary

• Base model: Supabase/gte-small
• Embedding dimension: 384
• Max sequence length: 512
• Checkpoint size (local): ~128MB
• Intended use: dense retrieval / semantic search

Benchmark (Retrieval-Only, No Generation)

Evaluated on fixed 150-case benchmark (50 queries each from mteb/fiqa, mteb/nfcorpus, mteb/scifact) using cosine similarity.

Aggregate Results

• hit@1: 0.6600
• hit@5: 0.8133
• mrr@10: 0.7311
• recall@10: 0.6212
• map@10: 0.5859

Comparison vs previous HF Neurosense

Previous Neurosense baseline (models/Neurosense):

• mrr@10: 0.7242

This compact checkpoint:

• mrr@10: 0.7311

Delta:

• +0.0069 MRR@10 on the same benchmark.

Usage

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("Sharjeelbaig/Neurosense-Compact-GTE-R1")
embeddings = model.encode([
    "what causes daytime sleepiness",
    "Sleep apnea can cause fragmented sleep and daytime fatigue"
], normalize_embeddings=True)

Notes

• This model is compact and practical for serverless constraints.
• It does not aim to replace very large embedding models globally; it is optimized for compact deployment with strong retrieval quality.