Fine-tuned MS-MARCO MiniLM-L-6 cross-encoder (22M params, English) for octomind capability auto-activation. Designed as the second stage after the muvon/octomind-embed bi-encoder retrieves the top-N candidates.

Trained with BinaryCrossEntropyLoss over (anchor, positive) and (anchor, hard_negative) pairs from the octomind-tap catalog. The hard negatives target the confusable neighbors the bi-encoder alone cannot reliably separate.

Wired into octomind via octolib's HuggingFace reranker provider (candle backend).