thinktecture/gemma3-1b-ft-nextera-f16

⚠️ 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

---

Gemma3-1B FT (f16) — Direct Answer + Tool-Result Synthesis (intent fallback)

Base model	google/gemma-3-1b-it (1.0B params)
License	Gemma Terms of Use — see MODEL_LICENSES.md
Training script	finetune/train_gemma3.py
Method	LoRA r=8, α=16, 7 epochs, lr=1e-4
Training data	data/training-data/gemma3_intent_{scenario}.jsonl + gemma3_synthesis_{scenario}.jsonl
Hardware tested	RTX PRO 6000 (CUDA), Apple M-series (Metal/MPS fallback)
Intended use	(1) Direct-answer responses (no-retrieval-needed questions, chitchat); (2) tool-result synthesis (turning JSON tool outputs into natural-language answers); (3) generative intent-classification fallback when the LogReg classifier isn't loaded. Primary intent path is the LogReg classifier — it handles ~93% of traffic deterministically in <25ms.
Out of scope	RAG synthesis from multi-document context (cross-contaminates facts — use Gemma3-4B FT for that). General-purpose chat. Multilingual beyond the scenario's training language.
Reference eval (Nextera)	Intent (fallback path): 96.7% / 174 of 180 (post-2026-05-15 retrain on corrected training data; pre-retrain was 93.3%). Direct-answer: 95%+. See docs/benchmarks/FINE_TUNING_INSIGHTS.md §10.
Known failure modes	On the fallback intent path: confuses tool_use vs rag_query when the question contains a number AND a topical noun (e.g. "What's the cost of GDPR compliance?"). LogReg primary handles these cleanly.