muvon/octomind-embed

Embedding model for octomind capability / skill auto-activation: ibm-granite/granite-embedding-30m-english (30M params, 6 layers, 384-dim, CLS-pooled, prefix-free, English) fine-tuned on trigger phrases from the octomind-tap capabilities + skills catalog and blended back into the base as a WiSE-FT model soup, which beats both the base and the raw fine-tune on the runtime gate (mean-of-top-3 cosine + threshold + margin).

Training: rule-based + LLM paraphrase augmentation, one epoch of CachedMultipleNegativesRankingLoss (scale 10) on in-class pairs and positive-aware hard-negative triplets, MatryoshkaLoss over [384, 256, 192, 128, 96], then weight interpolation with the base.

Files

• model.safetensors + 1_Pooling/ — sentence-transformers layout (fp32).
• onnx/model.onnx — fp32 graph.
• onnx/model_quantized.onnx — int8 (weight-only, per-channel, plain 8-bit range); this is what the octomind runtime loads. Pool with CLS as declared in 1_Pooling/config.json.

Use

octomind loads onnx:muvon/octomind-embed via octolib's ONNX provider (MODEL_NAME in octomind/src/embeddings/mod.rs). Runtime thresholds are model-specific and calibrated against the int8 graph (AUTO_ACTIVATE_THRESHOLD / _MARGIN in capability.rs, SEMANTIC_* in skill.rs).