yuvit-batra/qwen2.5-coder-7b-cadquery-mlx-4bit

Qwen2.5-Coder-7B CadQuery (MLX 4-bit)

A fine-tune of Qwen2.5-Coder-7B-Instruct that turns natural-language part descriptions into executable CadQuery (Python) scripts producing valid, dimensionally-correct B-rep solids exportable to STEP/STL.

Trained with LoRA (rank 16, 16 layers) via mlx-lm on an Apple M4 Max; this repo contains the fused 4-bit MLX weights. A GGUF build for llama.cpp/Ollama is published alongside this repo.

Prompt contract

Use this exact system prompt (the model was trained with it):

You are a CAD design assistant. Given a description of a mechanical part, respond with a complete CadQuery (Python) script that builds it. Use millimeters. The script must import cadquery as cq and assign the final single-solid model to a variable named result. Respond with only the code.

Set the stop token to <|im_end|>.

mlx_lm.generate --model yuvit-batra/qwen2.5-coder-7b-cadquery-mlx-4bit \
  --prompt "Design a 60mm diameter flange, 10mm thick, with a 20mm bore and six 6mm bolt holes on a 44mm circle." \
  --max-tokens 1024 --temp 0.2

Training data

12,477 validated text→CadQuery pairs (dataset v1.2):

• 8,468 synthetic samples across 22 parametric part families (plates, brackets, flanges, gears, shafts, enclosures, turned/chess-piece-like parts, phone stands, trays, hooks, clips, ISO fasteners...). Every sample was executed in a sandbox and verified to produce a valid single solid with a bounding box matching its description.
• 4,000 real-world human-designed models: DeepCAD construction sequences with Text2CAD annotations (CC-BY-NC-SA-4.0), converted to CadQuery and re-validated.
• 9 permissively-licensed GitHub CadQuery examples.

Evaluation (geometric, not textual)

Generated code is executed; solids are checked for validity and dimensional accuracy vs reference geometry (sorted-bbox within 10%/axis, volume within 15%).

Metric (n=25 test / n=33 novel)	Base Qwen2.5-Coder-7B	This model
Test: executes	28%	80%
Test: valid single solid	24%	80%
Test: bbox accuracy	33%	45%
Novel prompts: executes	64%	73%
Novel prompts: valid solid	48%	67%

The novel-prompt set includes chess pieces, phone docks, organizers, and other consumer parts phrased unlike the training templates.

Limitations & license

• Trained for single-part generation (no assemblies) in millimeters.
• ~32% of training data (the Text2CAD portion) uses normalized coordinates ("units"); prompts phrased in mm behave best.
• Dataset lineage includes CC-BY-NC-SA-4.0 data (Text2CAD annotations), so this model is released under CC-BY-NC-SA-4.0 — non-commercial use only.
• Validate generated geometry before manufacturing; the model can produce dimensionally plausible but incorrect parts.