BOOMI Stage-A 7B β€” text β†’ floor-plan SPEC (LoRA)

The language half of BOOMI: a LoRA adapter on Qwen2.5-Coder-7B that turns a natural-language home description into a structured, coordinate-free SPEC (room program, counts, target areas, adjacency, BHK, plot). A separate CP-SAT constraint solver (not learned) then turns that SPEC into valid, editable CAD geometry β€” see the Space.

This is the flagship, best-quality Stage-A model. For a CPU-friendly demo there is a smaller distilled variant; this 7B is for running on a GPU.

Intended use

  • Draft an editable residential floor-plan SPEC from a prompt, for human refinement.
  • Out of scope: sanction/construction use without licensed review; 3D/BIM; non-residential.

How to use

import torch
from peft import AutoPeftModelForCausalLM
from transformers import AutoTokenizer

tok = AutoTokenizer.from_pretrained("BDivyesh/boomi-stage-a-7b")
model = AutoPeftModelForCausalLM.from_pretrained(
    "BDivyesh/boomi-stage-a-7b", torch_dtype=torch.float16, device_map="cuda")
# Prompt format: see the BOOMI pipeline (stage_a_common.build_messages).

Output is a JSON SPEC; feed it to the BOOMI CP-SAT solver to get the DXF. Grammar- constrained decoding gives ~100% valid JSON.

Training

  • Base: Qwen/Qwen2.5-Coder-7B (Apache-2.0), LoRA r32/Ξ±64 (attn+MLP), completion-only SFT.
  • Data: ResPlan (MIT, ~14.9k vector residential plans) + RERA-derived Indian per-dwelling rows (Indian public building plans, PII-stripped, geometry-only) + deterministic-grammar synthetic captions. Fully permissive provenance.

Evaluation (Indian held-out)

  • Descriptive prompts (L1/L2/L3): 100% program-multiset-exact / Jaccard 1.00 / BHK 100% / valid-JSON 100%.
  • vs the ResPlan-only baseline: +75pp on Indian, no catastrophic forgetting (βˆ’1pp on ResPlan).
  • Terse one-line prompts are one-to-many (use a gallery / recall@K, not exact match).

Limitations

  • Terse prompts are ambiguous; the model may add a sensible default (e.g. a balcony).
  • On under-specified prompts it can emit BATH where Indian convention says TOILET (vocab carry-over from ResPlan) β€” handle with a deploy-time vocab map.
  • Drafts only; not sanctioned designs.

License: Apache-2.0 (base + adapter). Stack: OR-Tools/ezdxf/shapely all permissive.

Downloads last month
25
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for BDivyesh/boomi-stage-a-7b

Base model

Qwen/Qwen2.5-7B
Adapter
(46)
this model

Space using BDivyesh/boomi-stage-a-7b 1