kylebrodeur/ microfactory-node:latest

1 month ago

The first fine-tune attempt. Trained on gemma-3-1b-it with rank 16 for 3 epochs on deterministic targets. It parroted the same settings template for every input — it memorized, it did not judge. 

ollama run kylebrodeur/microfactory-node

Details

1 month ago

46487ddd4736 · 5.3GB ·

model
gemma4
·
7.46B
·
Q4_K_M
system
You are the Chief Engineer of a 3D-printing microfactory node — a small, fine-tuned local model tr
template
{{ if .System }}<start_of_turn>system {{ .System }}<end_of_turn> {{ end }}{{ if .Prompt }}<start_of_
params
{ "num_ctx": 4096, "stop": [ "<start_of_turn>user", "<end_of_turn>" ],

Readme

Microfactory Node: 3D Printer (LoRA v1 — historical)

⚠️ Do not use this model. It is a historical artifact kept for the paper trail.

The first fine-tune attempt. Trained on gemma-3-1b-it with rank 16 for 3 epochs on deterministic targets. It parroted the same settings template for every input — it memorized, it did not judge.

What went wrong

  • High rank (r=16) + many epochs (3) + deterministic targets = parrot
  • Base model (gemma-3-1b) was too small for the task

Lessons learned (applied in v2/v3)

  1. Noisy targets force judgment — temperature=0.7 during dataset generation prevents template memorization
  2. Low rank, single epoch — r=4 for 1 epoch: less capacity, less memorization, more attention to the actual job
  3. Larger base model — gemma-4-E4B-it (~4B effective) instead of gemma-3-1b

Use these instead

  • Recommended: ollama run kylebrodeur/microfactory-node-v3-qat (QAT, highest quality)
  • Alternative: ollama run kylebrodeur/microfactory-node-v2 (non-QAT, solid)

Links

License

Gemma license (inherited from base model).