scadcoder-qwen3.5-9b-3ep
A LoRA fine-tune of Qwen/Qwen3.5-9B for generating
parametric OpenSCAD models of bio-lab hardware, built to run on the Tiiny AI Pocket Lab.
- Base:
Qwen/Qwen3.5-9B — architecture Qwen3_5ForConditionalGeneration. The vision
encoder is frozen and preserved; only the language layers were adapted, then merged back to
bf16 safetensors (775 tensors, identical names to the base).
- Training: text-only LoRA (rank 32, dropout 0.1), 3 epochs, on 7,238 verified OpenSCAD
design examples (all kept at a 4096-token context length).
- Intended use: turning a natural-language spec into renderable OpenSCAD for lab parts
(racks, funnels, adapters, tube holders, etc.). Run with reasoning/thinking off.
The 3-epoch iteration of scadcoder-qwen3.5-9b. Research artifact from the "Prompt to Bench" project.