HistCAD: Constraint-Aware Parametric CAD Dataset

• HistCAD is a multimodal dataset providing a flat, history-based parametric sequence with synchronized modalities to capture design intent.
• It ensures constraint compliance with explicit encoding of ten geometric relationships, supporting native CAD editability and execution.
• HistCAD integrates academic and industrial modeling data to benchmark text-driven CAD generation with robust, intent-aware annotations.

Searching arXiv for HistCAD and closely related CAD dataset papers to ground the article with current citations. HistCAD is a large-scale, multimodal dataset for generative computer-aided design in which parametric modeling history, explicit geometric constraints, native CAD compatibility, and text supervision are aligned at the instance level. It was introduced to support editable, constraint-compliant, semantically enriched CAD generation by pairing a flat, history-based sequence representation with five synchronized modalities—modeling sequences, multi-view renderings, STEP-format B-reps, native parametric files, and textual annotations—and by adding an LLM-based annotation pipeline, AM$_\text{HistCAD}$, that produces descriptions of modeling process, geometric structure, and functional type (Dong et al., 8 Dec 2025).

1. Problem setting and design rationale

HistCAD was proposed against a specific diagnosis of prior CAD datasets. The paper argues that existing resources often omit explicit geometric constraints, lack industrial realism, provide weak semantic supervision, and encode modeling histories in structurally complex hierarchical forms that are difficult for sequence models to learn robustly (Dong et al., 8 Dec 2025). In that framing, the central problem is not merely representing final shape, but representing design intent in a form that remains editable and executable in native CAD environments.

This focus distinguishes HistCAD from datasets that privilege either geometry alone or procedural history alone. The dataset is explicitly designed for editable, constraint-compliant, semantically enriched CAD generation, meaning that downstream systems are expected not only to reconstruct a shape, but also to preserve the relational structure that makes parametric editing meaningful. A plausible implication is that HistCAD treats design intent as a first-class supervision target rather than as an incidental by-product of geometry.

The paper also positions HistCAD as a benchmark for text-driven CAD generation. That objective explains why the dataset couples procedural structure with richer annotations rather than relying only on geometric captions. In this formulation, the dataset is simultaneously a modeling corpus, a CAD-execution substrate, and a multimodal supervision resource.

2. Modalities, subsets, and corpus construction

HistCAD provides five aligned modalities for each model.

Modality	Role
Modeling sequences	Flat, history-based parametric representation
Multi-view renderings	Vision-language and image-based access
STEP-format B-reps	Standard geometric and topological exchange format
Native parametric files	Preservation of actual CAD history and software compatibility
Textual annotations	Descriptions of process, structure, and function

The dataset is divided into HistCAD-Academic and HistCAD-Industrial, with 160,501 high-quality modeling sequences in total: 152,360 fully aligned, constraint-aware modeling sequences in the academic subset and 8,141 professionally authored Siemens NX PRT models in the industrial subset (Dong et al., 8 Dec 2025). The academic portion is assembled from DeepCAD, SketchGraphs, and Fusion 360 Gallery. The industrial portion contributes authentic Siemens NX models with full parametric histories, sketches, feature definitions, and constraint relations, and it is emphasized as the source of more complex industrial modeling behavior, including rotated extrusions.

The academic pipeline is three-stage. First, DeepCAD models are decomposed into primitive sketch entities by face-loop decomposition. Second, hierarchical structures are flattened by symmetric difference:

$$\mathcal{P} = \bigtriangleup_{i=1}^n \left( \bigtriangleup_{L_{ij} \in \partial f_i} L_{ij} \right),$$

where $\mathcal{P}$ is the resulting set of non-overlapping primitives, $f_i$ are faces, $L_{ij}$ are boundary loops, and $\partial f_i$ denotes loops associated with face $f_i$ (Dong et al., 8 Dec 2025). Third, these primitives are aligned to SketchGraphs entities using geometric similarity and spatial proximity, with constraint propagation and pruning of redundant or conflicting constraints. From Fusion 360 Gallery, the pipeline recovers sketch-extrusion correspondences and produces 7,654 usable modeling sequences.

The resulting corpus is not a simple aggregation of pre-existing data. It is a normalized representation layer intended to reconcile heterogeneous sources into a single constraint-aware sequence format. This suggests that HistCAD is as much a data-integration framework as it is a dataset.

3. Sequence formalism, constraints, and CAD executability

The core HistCAD representation is a flat, history-based, parametric sequence built around sketch-and-extrude operations (Dong et al., 8 Dec 2025). Each sequence includes sketch plane definition, geometric primitives, geometric constraints, extrusion parameters, boolean operations, and rotated extrusion when applicable. The sketch plane is defined by a translation vector and Euler angles, allowing arbitrary placement in 3D space.

HistCAD supports three primitive types: lines, circles, and arcs. They are stored in compact parametric form as line endpoints, circle center and radius, and arc start, mid, and end points. A defining design choice is that primitives are stored as unordered sets, not as hierarchical loop structures; loop and face topology are inferred automatically from connectivity. The paper argues that removing explicit loop nesting simplifies the representation and makes it more robust for sequence learning while preserving geometry and editability.

A central technical contribution is explicit encoding of ten geometric constraints: coincident, parallel, perpendicular, horizontal, vertical, tangent, equal, concentric, fix, and normal (Dong et al., 8 Dec 2025). In the paper’s terminology, a sketch is geometrically constrained when primitives are bound by such relational rules, and a sequence is constraint-aware when those relations are explicitly stored for downstream reasoning and reconstruction. The significance is operational: if a circle radius changes, tangent lines update accordingly; if a coincident point moves, attached elements move too.

Extrusions are encoded by direction and length. HistCAD also supports rotated extrusions using a rotation axis, position and orientation, and start/end angles. Boolean operations include create new body, join, subtract, and intersect. The inclusion of these operations is what allows the representation to span both single-part and assembly-level modeling.

Executability is an explicit design goal. To instantiate models, the system reconstructs loops and faces from primitives using rule-based algorithms and converts each modeling command into geometric-kernel API calls (Dong et al., 8 Dec 2025). The paper states that the format is directly compatible with commercial CAD tools such as FreeCAD and SolidWorks. This is the basis for the claim that HistCAD supports native parametric editability rather than merely approximate regeneration of mesh geometry.

4. AM$_\text{HistCAD}$ and intent-aware semantic annotation

AM$_\text{HistCAD}$ is the dataset’s annotation module for generating richer text supervision directly from modeling history (Dong et al., 8 Dec 2025). Its purpose is to overcome the paper’s stated limitation of shallow captions that describe only coarse geometric appearance. The module targets three output types: modeling process, geometric structure, and functional type. It also uses natural language transcriptions (NLTs) as an intermediate representation.

The module operates in three stages. First, it parses modeling history. For each part, it computes 2D loops from the sketch, sorts them by area, and classifies them into outer contours and inner holes. It also computes a global 3D oriented bounding box for each part and uses the Separating Axis Theorem (SAT) on OBBs to classify pairwise inter-part relations as separate, touch, intersect, contain, or contained. Directional labels are inferred from centroid offsets.

Second, these parsed features are converted into NLTs. The transcriptions encode sketch primitives, loop and hole structures, geometric constraints, extrusion operations, coordinate transformations, inter-part spatial relations, and boolean operations (Dong et al., 8 Dec 2025). NLTs are not the final captions; they are structured prompts intended to stabilize reasoning over modeling history.

Third, a reasoning-capable LLM, specifically Qwen3-32B, generates the final annotations. The modeling-process prompt requests a concise paragraph describing sketch entities and parameters, constraints, extrusion direction and depth, coordinate transformations, and boolean combinations. The geometric-structure prompt focuses on visible geometry such as overall shape, holes, slots, cutouts, symmetry, repetition, and internal/external structure, explicitly excluding function and modeling steps. The functional-type prompt asks for a concise noun phrase such as hex head bolt, flat washer, or spur gear.

The paper distinguishes two annotation regimes in experiments. HistCAD$_H$ uses HistCAD’s four intent-aware annotations—modeling process, geometric structure, functional type, and natural language transcription—whereas HistCAD$$\mathcal{P} = \bigtriangleup_{i=1}^n \left( \bigtriangleup_{L_{ij} \in \partial f_i} L_{ij} \right),$$0 uses Text2CAD-style annotations (Dong et al., 8 Dec 2025). This split is important because it isolates the effect of richer semantic supervision from the effect of the sequence representation itself.

5. Compactness, editability, and benchmark outcomes

A major empirical claim of HistCAD is that flattening the modeling history reduces syntactic complexity and improves robustness. On a shared subset of 129,053 samples, the average token counts are reported as 1878.65 for DeepCAD, 560.24 for Text2CAD, 358.87 for HistCAD without constraints, and 476.11 for HistCAD with constraints (Dong et al., 8 Dec 2025). Over the same comparison, the average Chamfer Distance values are 7.76, 10.02, 7.59, and 7.59, respectively. The paper interprets this as evidence that HistCAD is substantially more compact than earlier sequence formats while maintaining geometric fidelity.

Subset-wise token statistics, measured with the Qwen3-0.6B tokenizer, further indicate differing complexity levels. HistCAD-DeepCAD has mean 711.56, median 361, and 95th percentile 1,963; HistCAD-Fusion360 has mean 691.77, median 414, and 95th percentile 2,104; HistCAD-Industrial has mean 1,348.27, median 631, and 95th percentile 4,098 (Dong et al., 8 Dec 2025). The industrial subset is therefore substantially more complex than the academic subsets.

The constraint distribution is also reported quantitatively: coincident 27.33%, horizontal 21.72%, perpendicular 16.97%, parallel 16.24%, vertical 7.57%, equal 4.11%, tangent 3.30%, concentric 2.54%, fix 0.19%, and normal 0.02% (Dong et al., 8 Dec 2025). This profile makes clear that constraint-awareness in HistCAD is not limited to a few rare annotations; it is structurally pervasive.

Editability is evaluated by importing models into FreeCAD and applying changes such as moving endpoints or changing arc radii. The reported result is that models without constraints often break tangency, equality, or concentricity, whereas models with HistCAD’s native constraints preserve these relationships and propagate edits correctly (Dong et al., 8 Dec 2025). This is one of the paper’s strongest arguments that explicit constraints materially affect downstream usability.

For text-driven generation, the paper fine-tunes Qwen3-8B with LoRA and reports that, under Text2CAD-style annotations, Text2CAD$$\mathcal{P} = \bigtriangleup_{i=1}^n \left( \bigtriangleup_{L_{ij} \in \partial f_i} L_{ij} \right),$$1 obtains IR 2.35%, Avg. CD 4.73, Med. CD 0.14; HistCAD$$\mathcal{P} = \bigtriangleup_{i=1}^n \left( \bigtriangleup_{L_{ij} \in \partial f_i} L_{ij} \right),$$2 (w/o c) obtains IR 1.48%, Avg. CD 3.30, Med. CD 0.13; and HistCAD$$\mathcal{P} = \bigtriangleup_{i=1}^n \left( \bigtriangleup_{L_{ij} \in \partial f_i} L_{ij} \right),$$3 obtains IR 1.40%, Avg. CD 3.00, Med. CD 0.13 (Dong et al., 8 Dec 2025). Here, IR denotes invalidity ratio, defined as the proportion of outputs that fail to compile into valid STEP files.

The effect of industrial parts is evaluated by training on HistCAD-DeepCAD, HistCAD-Academic, and full HistCAD. The reported weighted-average results are IR 6.76 / CD 25.75, IR 7.04 / CD 18.74, and IR 6.42 / CD 12.73, respectively (Dong et al., 8 Dec 2025). On the HistCAD-Industrial test set specifically, training on full HistCAD reduces IR from 17.36 to 11.40 and CD from 115.75 to 36.78. The paper treats this as evidence that industrial data improve generalization to real-world parts without harming academic-subset performance.

6. Scope boundaries, related usages, and significance

HistCAD, in the sense of (Dong et al., 8 Dec 2025), belongs to the domain of parametric computer-aided design, not medical computer-aided diagnosis. This distinction matters because the acronym appears differently elsewhere in the accompanying literature. In oral histopathology, the OCDC dataset is described as supporting histology-based computer-aided diagnosis (HistCAD) for OSCC tumor segmentation with 1,020 H&E-stained 640 × 640 patches and corresponding pixel-level masks (Santos et al., 2023). In coronary imaging, Coronary-GAN addresses virtual histology staining from OCT to H&E-like images under structural constraints (Li et al., 2022). In breast pathology, a compact texture-oriented CNN is presented for benign-versus-malignant histopathological image classification on BreaKHis (Matos et al., 2019). These works are adjacent in acronymic usage but belong to a different technical lineage.

Within CAD modeling proper, a common misconception is to treat HistCAD as merely a larger shape corpus. The paper’s actual claim is narrower and more technical: the value lies in explicit constraints, flattened sequence format, multi-type annotations, and native-file alignment, all of which are intended to improve robustness, parametric editability, and text-driven generation (Dong et al., 8 Dec 2025). Another misconception is that semantic supervision in CAD datasets is exhausted by geometric captions. HistCAD rejects that assumption by separating modeling process, geometric structure, and functional type into distinct annotation targets.

Taken together, HistCAD defines a particular view of generative CAD benchmarking: geometry alone is insufficient, procedural history alone is insufficient, and text alone is insufficient unless these are aligned through constraint-aware parametric structure. Its broader significance is therefore methodological. It proposes that editable CAD generation should be evaluated not only by geometric similarity, but also by whether generated artifacts remain executable, modifiable, and semantically interpretable in actual CAD workflows (Dong et al., 8 Dec 2025).