Training data
8,784 single-turn examples, 3 epochs, LoRA rank 32 (dropout 0.1) on the language layers:
- the 7,238 render-verified examples of v2 (
scadcoder-qwen3.5-9b-3ep): 34 parametric part families plus repair examples;
- 336 examples distilled from Claude Opus on bench-style part descriptions, each render-verified (watertight, bounding box within 15% of the specification), weighted ×3;
- 269 of the model's own rejection-sampled generations that passed the verifier, weighted ×2.
Evaluation caveat. Several of the distilled part families follow the wording of Prompt-to-Bench task prompts
(e.g. tube racks, funnels, gel combs, adapter sleeves) with different dimensions, so this model's scores on those
tasks are in-family, not held-out transfer. v2 is the held-out baseline. Details:
https://github.com/ATinyGreenCell/prompt-to-bench (§6.4).