vykrum/hywe-training-data

---

HYWE Architectural Training Data

A highly structured, synthetic dataset of procedural architectural programming, design intent, and topological layouts. This dataset is generated client-side by the HYWE Core Engine and curated via the Hynteract serverless pipeline to provide a robust, logic-driven substrate for machine learning models in architectural design (AEC) and spatial reasoning.

• Ecosystem Repository: github.com/vykrum/Hywe
• Interactive Sandbox: hywe.in

---

Technical Essence & Philosophy

Traditional generative ML models for architecture rely on heavy boundary representation formats (like OBJ, IFC, or DXF) or unstructured pixel matrices, which fail to capture functional spatial relationships.

HYWE (Hygrid Woven Ensemble) models design space as a function of discrete computational logic. It rejects manual drafting and continuous geometric solvers in favor of a hybrid orthogonal-hexagonal grid (Hygrid). In this system, spatial adjacency is a direct mathematical consequence of defined architectural programming constraints rather than absolute geometric coordinates.

This dataset provides the bridging data to train AI models that can generate deterministic topologies directly from linguistic design narratives.

---

System Architecture Flow

Designer Intent ➔ HYWE Syntax ➔ Deterministic Topology ➔ Spatial Configuration ➔ Hynteract Structuring ➔ JSONL Dataset ➔ AI Training

---

Data Curation & Architecture Flow

The data is captured through a closed-loop computational pipeline where the deterministic geometry logic of HYWE remains isolated from the probabilistic data pipeline of Hynteract.

graph TD
    A1[Interactive Node Tree Input] --> B[HYWE Syntax]
    A2[Interactive Boundary Editor] --> B[HYWE Syntax]
    B --> C(Lexel: Architectural Programming and Flow Parsing)
    C --> D(Hexel: Atomic Spatial Primitive)
    D --> E(Coxel: Simultaneously Evolving Hexel Clusters)
    E --> F(Xyxel: Coxel Configuration and Planar Layout)
    F --> G(Zaxel: Xyxel Stacking and Volumetric Massing)
    
    F --> F1[SVG Rendering]
    G --> G1[WebGPU Massing]
    
    F --> H[Spatial Analysis]
    F --> I[Batch Processing]
    
    I -.-> DatasetLabel((Hynteract: AI Dataset))
    DesignIntent[Design Intent Narrative] --> DatasetLabel

1. Procedural Variation: The designer builds or batch-generates structural arrangements within HYWE.
2. Topological Compression: HYWE encodes the hierarchical spatial configurations into an ultra-dense Base36 alphanumeric token (HYWE Syntax).
3. Pipeline Commitment: The Hynteract Vercel API endpoint securely captures the Base36 token, pairs it with the natural language description, and commits structured JSON Lines (.jsonl) files directly to this Hugging Face repository.

---

The Computational Generation Pipeline

For a deep dive into the engine-specific terminology used below (Lexel, Hexel, Coxel, etc.), refer to the HYWE Architecture Wiki.

Every spatial layout committed to this dataset is resolved using a zero-dependency, first-principles geometric compiler. Rather than relying on heuristic geometric optimization or probabilistic CAD models, the HYWE Core Engine compiles layout configurations deterministically through five functional stages:

1. Lexel (Linguistic & Flow Parsing):
   • Role: Translates hierarchical trees and flow connections (defining circulation routes and adjacency rules) into active logical constraints.
2. Hexel (Atomic Priming):
   • Role: Maps discrete space coordinates onto a hybrid orthogonal-hexagonal spatial lattice (Hygrid) using integer arithmetic, preventing floating-point coordinate drift.
3. Coxel (Cluster Growth):
   • Role: Groups discrete hexels into simultaneously expanding clusters to represent coherent programmatic zones, nodes, or functional areas.
4. Xyxel (Planar Subdivision):
   • Role: Subdivides, proportions, and fits the growing coxel boundaries within irregularity constraints to resolve a finalized 2D planar floor layout.
5. Zaxel (Volumetric Stacking):
   • Role: Stacks planar layout matrices vertically to distribute programmatic nodes across multiple levels, resolving vertical circulation and massing constraints.

---

Data Schema

Each record is stored as a .jsonl file in the data/ directory. The top-level structure per record:

Field	Type	Description
definition	string	The single source of truth HYWE Syntax representation containing design constraints, block attributes, boundaries, and cell specs.
description	string	A natural language spatial narrative detailing the designer's intent, scale, flow, and typology.
configuration	string[]	An array of container-scoped geometry strings, each encoding exactly 24 layout string variations in a Base36-compressed format.

Complete Layout Sweeps: Although the definition field contains a default sequence rule in its Q attribute (representing the active configuration at the time of export), the dataset itself is not restricted to that single configuration. The accompanying configuration field contains the resolved layout strings for all 24 possible sequence permutations, enabling comprehensive training across all sequence rules.

Configuration Layout String Format

Each string in configuration encodes one architectural container (Level or Nest) along with all its 24 sweeps:

Marker(ID_1;ID_2;...;ID_N | Variation_0 | Variation_1 | ... | Variation_23)

• Marker: L0, L1, N1, N2, etc.
• IDs: Semicolon-separated list of prefix-stripped local IDs (e.g. 1;1.1;1.2).
• Variations: 24 blocks separated by |. Each block contains semicolon-separated node coordinates.
• Coordinates: Flat, comma-separated Base36 X,Y pairs per hexel (e.g. F,5,H,5 = cells at (15,5) and (17,5)).
• Positional Alignment: The coordinate groups within each variation block are strictly positionally aligned with the IDs list in the header. (e.g., the 3rd coordinate string in Variation 5 belongs to the 3rd node listed in the header).

Sequence Mappings (Indices 0 to 23)

The index of each layout string variation block corresponds to the following sequence rule:

Index	Sequence Name	Index	Sequence Name		Index	Sequence Name	Index	Sequence Name
0	VRCWEE	6	VRCWWW		12	HRCWNN	18	HRCWSS
1	VRCCEE	7	VRCCWW		13	HRCCNN	19	HRCCSS
2	VRCWSE	8	VRCWNW		14	HRCWNE	20	HRCWSW
3	VRCCSE	9	VRCCNW		15	HRCCNE	21	HRCCSW
4	VRCWSW	10	VRCWNE		16	HRCWSE	22	HRCWNW
5	VRCCSW	11	VRCCNE		17	HRCCSE	23	HRCCNW

---

Citation

If you use this dataset or engine in your research or projects, please cite it using the following BibTeX format:

@misc{hywe2026dataset,
  author = {Subbaiah, Vikram},
  title = {HYWE Architectural Training Data: A Structured Dataset of Procedural Architectural Programming and Topological Layouts},
  year = {2026},
  publisher = {Hugging Face},
  journal = {Hugging Face Repository},
  howpublished = {\url{https://huggingface.co/datasets/vykrum/hywe-training-data}}
}

---

License

To align with the open-science principles of this project while protecting the underlying source code, the HYWE ecosystem utilizes a dual-licensing structure:

• This Dataset Card and all Datasets inside this Hugging Face repository are licensed under the Creative Commons Attribution 4.0 International (CC BY 4.0) license. You are free to share and adapt this data, provided you give appropriate credit to the author.
• The HYWE Core Engine source code (the F# compiler, frontend, and coordinate solvers) is licensed under the highly permissive MIT License.