config	records	given	answer
`recode`	953,081	an eight-view render of a solid, and nothing else	the CadQuery program that builds it
`comprehension`	953,081	the program as text, no image	a property of the solid it builds

from datasets import load_dataset
r = load_dataset("AI4Manufacturing/w_cad2", "recode",        split="train")
c = load_dataset("AI4Manufacturing/w_cad2", "comprehension", split="train")

The source has no model input at all

This is the point of the release. CAD-Recode ships programs and nothing else — its own metadata says has_input: false. There are no images and no stored point clouds, so the task it names, recover the program, cannot actually be posed from what ships. Every program here was executed and the resulting solid rendered; that is what makes the question answerable.

Two tracks, deliberately sharing no evidence: recode carries an image and a null cadquery column (the program is the answer there), comprehension carries the program and no image. A record that offered both would let a model read the answer instead of looking.

⚠️ Exact code match is not a valid score

The same shape has many valid programs. Scoring a prediction by string comparison against annot will mark correct reconstructions wrong. Execute the prediction and compare geometry (Chamfer distance / IoU). This warning is repeated in every record's metadata.eval_note.

Every render is varied, per record, from a seed

Fixed camera angles, lighting and colour are a format a model can learn instead of the geometry, and nothing learned that way transfers to a CAD screenshot taken elsewhere. So each record's render varies: view direction by up to 12° about a random axis, camera roll up to 8°, key light up to 18°, plus material grey, ambient level, framing slack and background.

Never varied, because they carry meaning rather than style: orthographic projection, the single framing shared by all eight views of a solid, and the number of views.

The seed is derived from the program text, so a rebuild is pixel-identical, and the values actually used are in metadata.jitter_applied — each record documents its own render. Measured over 200,000 consecutive records: 200,000 distinct seeds.

Renders are produced by a deterministic software rasteriser (no GPU, no driver, no sampling), so the same program gives the same pixels on any machine.

Know what you are training on

The generator unions random primitives and nothing forces them to touch, so these are frequently not single connected parts:

bodies in one "solid"	share
1 — a single connected part	42.9%
2	35.8%
3	16.1%
4 or more	5.3%

Most remain legible — the median solid covers 23% of the frame — but 40,365 records (4.2%) are flagged likely_unreadable (footprint under 2% of frame, or more than three loose bodies). A shape too small or scattered to see cannot determine a program; filter on the flag if that matters to you.

Deduplicated, and what is still flagged

3.03% of the source's train programs are exact duplicates — 29,766 extra copies, one program repeated 60 times, which over-weights those samples in a generative task. This release ships the deduplicated set: 952,099 distinct train programs + all 982 val = 953,081. Every record has duplicate_rank == 0; the field is kept so the filter is auditable.

Still flagged rather than removed:

flag	meaning
`also_in_train`	35 of the 982 val programs appear verbatim in train. val is a same-distribution holdout from the same generator — a training monitor, never a benchmark — and this part of it is leaked. The source's own evaluation is done on external datasets (DeepCAD / Fusion360 / CC3D).
`likely_unreadable`	footprint < 2% of frame, or > 3 disconnected bodies
`footprint_mean`	share of the frame the solid covers, averaged over the eight views
`n_bodies`, `n_faces`, `bbox_dims`	measured from the executed solid

The `r` convention, if you execute predictions

The final solid must be bound to a variable named r. That is not our rule — every official CAD-Recode consumer reads globals()['r'].val(), and their conversion code swallows failures with a bare except: pass, so a program binding anything else does not error, it silently produces nothing. Our executor surfaces it instead. The prompt states the requirement explicitly.

Roles

Roles: annot is the machine-parseable gold and, for now, also the SFT target — the default view is query (+ image for recode, + cadquery for comprehension) → annot. reasoning is null by design: chain-of-thought is a separate annotation stage. In recode the answer is a program, so it is a generation target scored geometrically (see above); in comprehension it is a short value (a count or three integers) and is exactly checkable. When a CoT layer is added later it becomes the imitation target and annot stays the gold — it is never overwritten.

Fields

query (from 32 hand-written paraphrases per track; the source had one) · image (recode only, 2078×1042, eight 512 px panels) · cadquery (comprehension only — null in recode, where it would be the answer) · annot · reasoning (null) · cate = E · task = T-E1 · metadata.

Splits. Everything ships as one train split; the source's split is preserved in metadata.split_official (952,099 train / 982 val). Both tracks derive from the same programs, so split by metadata.model_id across tracks if you use both, or a program seen in one will leak into the other's evaluation.

Provenance

Built by forge_model/CAD_RECODE/cadrecode_build.py --dedup on the forge_cad toolkit (forge_agent/forge_cad), verified by verify_cadrecode_build.py, published by publish_w_cad2.py. No model API was used — the gold is the source's own program, and every derived property is computed by executing it. Verification re-executes a sample of the gold and recomputes the properties from the fresh solid: 236/236 executed, 236/236 properties matched.

Derived from CAD-Recode, CC BY-NC 4.0 (non-commercial). Please cite the original:

Rukhovich, D., Cherenkova, K., Kacem, A., Aouada, D. et al. CAD-Recode: Reverse Engineering CAD Code from Point Clouds.

Downloads last month: 11

Total file size:

44.7 GB

You need to agree to share your contact information to access this dataset

w_cad2 — CAD-Recode as VLM data