muvon/octomind-embed

Fine-tuned BGE-small-en-v1.5 (33M params, 384-dim) for octomind capability auto-activation.

Trained on trigger phrases from the octomind-tap capabilities + skills catalog with rule-based + LLM paraphrase augmentation, using MultipleNegativesRankingLoss on both in-class pairs and hard-negative triplets mined from confusable neighboring labels.

Use

Wired into octomind via octolib's HuggingFace embedding provider (candle backend). Set MODEL_NAME in octomind/src/embeddings/mod.rs to muvon/octomind-embed.

Paired reranker

muvon/octomind-rerank is the second-stage cross-encoder trained on the same hard-negative data.