BDivyesh/boomi-stage-a-1.5b

BOOMI Stage-A 1.5B — text → floor-plan SPEC (CPU-friendly LoRA)

The efficient Stage-A model for BOOMI: a LoRA adapter on Qwen2.5-Coder-1.5B that turns a natural-language home description into a structured SPEC (rooms, counts, areas, adjacency, BHK, plot). Small enough to run on a CPU — it powers the live demo Space. A separate CP-SAT solver turns the SPEC into valid, editable CAD geometry.

This is the small/efficient variant. On the narrow text→SPEC task it matches the flagship 7B (BDivyesh/boomi-stage-a-7b): held-out Indian eval 75% program-multiset-exact / 100% valid-JSON / 100% BHK / Jaccard 0.85 (vs 7B: 75% / 100% / 100% / 0.86) — because the task is a fixed-schema parse and grammar- constrained decoding guarantees valid JSON regardless of size.

Use

import torch
from peft import AutoPeftModelForCausalLM
from transformers import AutoTokenizer
tok = AutoTokenizer.from_pretrained("BDivyesh/boomi-stage-a-1.5b")
model = AutoPeftModelForCausalLM.from_pretrained("BDivyesh/boomi-stage-a-1.5b", torch_dtype=torch.float32)
# Prompt format: BOOMI stage_a_common.build_messages. Output JSON -> CP-SAT solver -> DXF.

Training

Base Qwen/Qwen2.5-Coder-1.5B (Apache-2.0), LoRA r32/α64, completion-only SFT on ResPlan (MIT) + RERA-derived Indian per-dwelling rows (PII-stripped, geometry-only) + grammar captions. Fully permissive provenance.

Limitations

Terse prompts are one-to-many (the model may add a sensible default); can emit BATH where Indian convention says TOILET on under-specified prompts. Drafts only — not for sanction/construction without licensed review. License: Apache-2.0.