Text Generation
PEFT
Safetensors
English
lora
qwen2.5
text-to-cad
floor-plan
dxf
architecture
india
conversational
Instructions to use BDivyesh/boomi-stage-a-7b with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- PEFT
How to use BDivyesh/boomi-stage-a-7b with PEFT:
from peft import PeftModel from transformers import AutoModelForCausalLM base_model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-Coder-7B") model = PeftModel.from_pretrained(base_model, "BDivyesh/boomi-stage-a-7b") - Notebooks
- Google Colab
- Kaggle
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 (β1pp 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
BATHwhere Indian convention saysTOILET(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