kylebrodeur/microfactory-node-gguf

Microfactory Node — Chief Engineer (GGUF)

Quantized GGUFs of three LoRA-fine-tuned variants of `google/gemma-4-e4b-it`, trained on real 3D-printer outcomes to predict where a print will fail and propose settings before the nozzle moves.

Both distribution paths point at the same blobs:

• —`ollama.com/kylebrodeur` — public Ollama registry, one-command pulls
• —`huggingface.co/kylebrodeur/microfactory-node-gguf` (this repo) — canonical GGUFs + template/system/params config

The QAT model was trained with simulated 4-bit quantization, so it retains more quality after quantization than the standard v2. Use q4_k_m for balanced quality/size, or q4_0 (the quant Google's QAT was trained for) for the highest fidelity reconstruction of the QAT model.

Run with Ollama (public registry — easiest)

# recommended
ollama run kylebrodeur/microfactory-node-v3-qat

# QAT-native quant
ollama run kylebrodeur/microfactory-node-v3-qat:q4_0

# other variants
ollama run kylebrodeur/microfactory-node-v2
ollama run kylebrodeur/microfactory-node

Run with Ollama (this HF repo — no download step)

Ollama can pull GGUFs directly from HF — the template, system, and params files in this repo configure the Gemma 4 chat template, the Chief Engineer system prompt, and tuned sampling automatically:

ollama run hf.co/kylebrodeur/microfactory-node-gguf:microfactory-node-v3-qat.gguf
ollama run hf.co/kylebrodeur/microfactory-node-gguf:microfactory-node-v3-qat-q4_0.gguf
ollama run hf.co/kylebrodeur/microfactory-node-gguf:microfactory-node-v2.gguf
ollama run hf.co/kylebrodeur/microfactory-node-gguf:microfactory-node.gguf

See the HF × Ollama docs for the hf.co/... URI form and how Ollama discovers the auxiliary config files.

Run with llama.cpp

hf download kylebrodeur/microfactory-node-gguf microfactory-node-v3-qat.gguf --local-dir .
llama-cli -m microfactory-node-v3-qat.gguf -p "PLA overhang at 22C, 45% humidity"

Use the live demo

The Hugging Face Space `build-small-hackathon/microfactory-lab` runs the full Chief Engineer UI against these adapters (ZeroGPU + a Modal-hosted OpenAI-compatible endpoint as fallback). Source repo: `kylebrodeur/microfactory-lab`.

The full conversion + publishing pipeline (LoRA → Modal merge → llama.cpp quantize → HF Hub → ollama.com) is documented in `learn/finetune/OLLAMA_PUBLISHING.md`.