tzachi2222/architect-ai-boomi

Architect AI Dataset

Trains an LLM to act as a residential architectural planner: given a Brief (what the user wants), Site (the physical plot), and a set of Constraints (explicit requirements, at varying levels of detail), produce an Architectural SPEC — a room program, zoning, and functional relationships. Geometry (coordinates, wall placement, doors) is intentionally out of scope for this model; that belongs to a separate geometry solver.

Source and provenance

Derived from BDivyesh/boomi-stage-a-text-spec (Apache-2.0) via a from-scratch preprocessing pipeline (src/ in the Architect AI repository). See docs/DATA_CONTRACT.md for the exact, verified semantics of every field (what target_area_m2 vs site.area_m2 mean, why relationships are type-level only, the area-tolerance policy, etc.) — every claim there was checked directly against the raw source data, not assumed.

Generation contract: one BOOMI source plan → exactly four variants ({plan_id}-C0, -C1, -C2, -C3), one per constraint-richness level. BOOMI's 4 caption-detail levels (L0-L3) share an identical underlying spec per plan (verified) and are collapsed to one canonical source row (L3 preferred) — they do not create additional training examples.

Splits

Deterministic 80/10/10 split by source plan, computed once over all unique BOOMI plans before any variant was generated, so all 4 difficulty-variants of a plan always land in the same split. Verified zero source_plan_id overlap between splits.

split	examples	unique source plans
train	54,044	13,511
validation	6,756	1,689
test	6,760	1,690

One plan (id 15861) was excluded entirely — a confirmed BOOMI source defect (degenerate site dimensions, plot.area_m2 = 0), not a processing error.

Difficulty levels (C0-C3)

Controls how many, and how varied, the input constraints are — not architectural quality:

level	constraints	intent
C0	0	Brief + Site only
C1	1-2	a couple of simple facts
C2	3-5	a moderate, type-diverse set
C3	6-8	a richer, type-diverse set — capped below the full extractable fact set

Constraint sampling is deterministic per (source_plan_id, difficulty) (seeded from GENERATION_SEED:plan_id:difficulty), so regenerating one variant always reproduces identical constraints.

Target-leakage measured, not assumed: even the single highest-coverage example never exposes more than ~59% of the extractable facts about its target (C0 mean 0.22 → C3 mean 0.38, max 0.59) — per-room areas and zone assignments are never exposed in the input at any difficulty, only room types/counts, some relationships, and total area.

Fields

• input.brief: what the user wants (building type, target built area, bedroom/bathroom/balcony counts, free-text preferences).
• input.site: the physical plot (width_m/length_m are bounding dimensions only — BOOMI sites are frequently non-rectangular, so area_m2 is authoritative and never derived from width×length).
• input.constraints: a sampled subset of facts the target actually satisfies, each tagged priority (HARD/SOFT) and source_type (USER_REQUIREMENT for room counts/total area, OBSERVED_GEOMETRY for relationships — an observed layout choice in one solution, not a universal rule the model must reproduce).
• target.program: room types, counts, area per room, and zone.
• target.zones: which room types belong to each zone (PUBLIC/PRIVATE/SERVICE/CIRCULATION/OUTDOOR).
• target.relationships: type-level adjacency/door/window connections (BOOMI has no room-instance IDs, so relationships cannot be instance-level — a documented source limitation, not a shortcut).
• target.circulation: room types serving circulation (e.g. staircases).

Known limitations (V1 scope)

• Single-floor residential only; BOOMI carries no floor count, so a small number of examples (108, ~0.16%) have built_area > site_area, plausible for undeclared multi-floor plans — kept, not removed, and worth revisiting before claiming multi-floor support.
• Rare room types: PARKING (1,336 examples, ~2%), STAIRCASE (2,716, ~4%) — usable but under-represented; no rebalancing has been applied.
• No Israeli (or any other) regulatory constraints are encoded — those belong to a separate, runtime-retrieved regulation engine by design, never baked into this dataset or model weights.

Not yet done

No base-model baseline evaluation, no fine-tuning.