thinktecture/gemma3-4b-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

---

Gemma3-4B FT (f16) — RAG Synthesis (+ Vision)

Base model	google/gemma-3-4b-it (4.3B params, multimodal: text + vision via mmproj)
License	Gemma Terms of Use
Training script	finetune/train_gemma3_4b.py
Method	LoRA r=16, α=32, 3 epochs, lr=5e-5
Training data	data/training-data/gemma3_4b_synthesis_{scenario}.jsonl (RAG passages + grounded answers)
Hardware tested	RTX PRO 6000 (CUDA). MPS works but slow; QLoRA via --qlora for ≤24GB VRAM
Intended use	RAG response synthesis — given retrieved passages and a user question, produce a grounded, source-faithful answer. The vision channel (mmproj) remains base-only.
Out of scope	Tool calling (delegated to Qwen3.5-4B FT). Free-form chat without retrieved context.
Reference eval (Nextera)	RAG keyword grounding: 96% on 25-query holdout. See docs/benchmarks/EVAL_RESULTS_*.md.
Known failure modes	Will occasionally synthesise across documents that share lexical overlap but different domains — mitigated by the rewrite-query step that pre-filters retrieval.