CAD-RLVR Engineering Guides

Practical documentation for integrating spacetime-ml/cad-cua and verify_step.py into PyTorch and Ray post-training pipelines.

Implementation Playbooks

Core Integration Guides

Deep dives into 3D CAD computer-use trajectories and deterministic B-Rep geometric rewards

Loading spacetime-ml/cad-cua from Hugging Face

How to load the default (62 episodes), models (48 submission zips, 215 CAD files), and benchmark (20 tasks) splits using Hugging Face Datasets and inspect synchronized 34,551 PNG frames and 260,203 telemetry events.

Topics covered:

default / models / benchmarkSHA256SUMS verification260,203 telemetry events34,551 lossless PNGs

Deterministic B-Rep Verification with verify_step.py

Run verify_step.py on candidate ISO-10303-21 .STEP outputs to compute composite RLVR reward R in [0, 1] across bounding-box IoU (35%), volume/area conservation (25%), B-Rep topology (20%), and bore radius spectrum (20%).

Topics covered:

MANIFOLD_SOLID_BREPCLOSED_SHELL >= 1Euler-Poincaré χR >= 0.85 pass gate

PyTorch & Ray Distributed CAD-RLVR Rollouts

Connect distributed GRPO and PPO rollout workers in Ray/PyTorch to parallel CAD execution environments (FreeCAD, OpenCASCADE, CadQuery, Build123d, SolidWorks, PTC Creo) with deterministic B-Rep reward feedback.

Topics covered:

Ray rollout workersGRPO / PPO rewardsHeadless CAD kernelsZero reward hacking

Training 6-DoF GUI Computer-Use Agents (CUA)

Best practices for behavior cloning and RL post-training on high-density desktop CAD trajectories featuring 3D viewport orbiting, feature-tree navigation, sketch constraint snapping, and parametric dimensioning.

Topics covered:

6-DoF viewport orbitSub-pixel sketch snapsMulti-view ISO-128 promptsapp_context_updated_at

Step-by-Step Workflow

CAD-RLVR Integration Path

From downloading the 15.48 GB public release to scaling distributed RL rollouts

1. Quickstart: Dataset Download & Integrity Check

Clone or stream https://huggingface.co/datasets/spacetime-ml/cad-cua (401 files, 15.48 GB, CC BY 4.0) and run `shasum -a 256 -c SHA256SUMS` to verify all episode archives, 48 ISO-128 blueprint pairs, and 48 model zips.

2. Evaluating Models on the 20-Task Blueprint Benchmark

Iterate over `data/benchmark.jsonl`, prompt your VLM/CUA or CAD-code agent with each ISO-128 4-view orthographic + isometric blueprint (`blueprints/cad-XXXXXX.png`), export a candidate `.step` solid, and score it with `python3 verify_step.py`.

3. Multi-Modal Trajectory Format & Action Space

Each episode in `data/train.jsonl` pairs 9-pass OCR privacy-scrubbed lossless PNG keyframes with timestamped mouse clicks, drags, keystrokes, 3D viewport scroll/orbit events, and verified active window metadata (`app_context_updated_at`).

4. Scaling Private Enterprise CAD-RLVR Clusters

Deploy thousands of parallel CAD-RLVR workers in isolated VPC or air-gapped enclaves with custom parametric assemblies and dense per-feature geometric reward shaping.

Need custom CAD-RLVR cluster integration?

Work with our systems engineers to integrate multi-suite 3D CAD environments and B-Rep verifiers into your training stack.

Partner With Us