thinktecture/intent-logreg-nextera

⚠️ Conference talk demo — not production weights.

This model accompanies a conference keynote on local on-device AI. Published as a reference for the fine-tuning patterns shown on stage — not a deployable artefact. No security audit, no SLA, pinned to the talk's state.

• Source repository: thinktecture-labs/local-multi-model-agent-slm
• Threat model + out-of-scope: SECURITY.md
• Licensing details: MODEL_LICENSES.md
• All five models in the stack: Collection — Local Multi-Model Agent — nextera fine-tunes

---

LogReg Intent Classifier

Base	scikit-learn LogisticRegression, multinomial, L2 penalty
License	Apache-2.0 (this repo) — but inputs are EmbeddingGemma vectors so the Gemma Terms cover the embedding step
Training script	training/train_intent_logreg.py
Method	LogReg on FT-EmbeddingGemma's 768-dim output vectors. Held-out 90/10 split. ~2 minutes on CPU.
Training data	Same as Gemma3-1B intent: data/training-data/gemma3_intent_{scenario}.jsonl (re-embedded with the FT EmbeddingGemma)
Hardware	CPU is sufficient. Requires the FT EmbeddingGemma llama-server running on port 9092/9096 to embed training examples.
Intended use	Replaces the 1B generative classifier as the primary intent router. ~10ms per query (vs ~200ms for the 1B). Same accuracy on the standard eval set.
Out of scope	Anything that requires generation (it's a 3-way classifier). Low-confidence predictions (< 0.60 threshold, configurable in intent_classifier_logreg.py) are overridden to direct_answer as a safe fallback intent. The 1B generative classifier is only used as a load-time fallback when the LogReg model file is absent, not as a per-query confidence fallback.
Reference eval (Nextera)	96.1% on 180-query eval set. ~10ms per classification (single CPU thread).
Known failure modes	When the EmbeddingGemma FT changes, the LogReg weights become invalid — intent_classifier_logreg.py:13-15 warns about this coupling. Re-train both together.