tap-model-training

What problem does it solve?

tap-model-training solves stale capability auto-activation by retraining and republishing the muvON/octomind-embed model whenever your capability triggers or skill semantic routing rules change, ensuring the runtime embedding gate stays aligned with your intent supervision.

Core Features & Use Cases

• When to retrain: Detects supervision-set changes caused by edits to capabilities//config.toml triggers and semantic(...) phrases in skills//SKILL.md, and distinguishes cases where retraining is unnecessary.
• End-to-end training + eval pipeline: Runs LLM augmentation for new intents, rebuilds datasets (pairs/triplets/holdout), performs iterative hard-negative mining and training with the CachedMNRL + GISTEmbed + Matryoshka loss stack, and validates with both holdout and eval_real gating.
• Publish with regression guardrails: Exports ONNX for ORT consumers and publishes to HuggingFace only after eval_gate regression checks, then updates the baseline to gate future publishes.
• Operational recovery and calibration: Provides checkpoint resume steps for interrupted runs and calibrates AUTO_ACTIVATE_THRESHOLD/MARGIN (capability.rs) and semantic thresholds/margins (skill.rs) from eval_real.

Quick Start

Run tap-model-training to retrain and republish the tap embedding model after you add or edit capability triggers or semantic(...) skill rules.