Choukrijer/buildwellai-qwen3-14b

BuildwellAI Qwen3-14B

Fine-tuned Qwen3-14B model for UK building regulations and construction calculations.

Model Description

This model has been fine-tuned on 50,885 examples covering:

• UK Building Regulations (Parts A-O)
• Thermal bridging and PSI values (SAP 10.2)
• U-value calculations
• Fire safety (Part B, BS 9991)
• Water efficiency (Part G)
• BREEAM, Passivhaus, WELL standards
• Structural calculations (Eurocodes)

Tool Calling

The model is trained to call MCP (Model Context Protocol) tools for calculations:

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("Choukrijer/buildwellai-qwen3-14b", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("Choukrijer/buildwellai-qwen3-14b")

messages = [
    {"role": "system", "content": "You are BuildwellAI, expert in UK building regulations."},
    {"role": "user", "content": "What is the PSI value for thermal bridge junction E1?"}
]

tools = [
    {"type": "function", "function": {"name": "get_psi_value", "description": "Get PSI value", "parameters": {"type": "object", "properties": {"junction_code": {"type": "string"}}, "required": ["junction_code"]}}}
]

text = tokenizer.apply_chat_template(messages, tools=tools, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
output = model.generate(**inputs, max_new_tokens=500)
print(tokenizer.decode(output[0]))

Available MCP Tools

The model can call 42 specialized MCP servers with 477 tools:

Category	MCPs
Thermal Performance	psi-thermal-bridge, thermal-break, condensation-glaser, wufi-hygrothermal
Energy Assessment	sap10, sbem, air-permeability, passivhaus
Building Regulations	structural-part-a, water-efficiency-part-g, overheating-part-o, ventilation-part-f
Fire Safety	fire-safety, smoke-ventilation, evacuation, cfd-fire-smoke
Sustainability	breeam, well, embodied-carbon, lca, biodiversity-net-gain
Daylighting	daylight-factor, adf-modelling, sunlight-overshadowing

Training Details

• Base model: Qwen/Qwen3-14B
• Training examples: 50,885
• Method: LoRA fine-tuning (rank 64, alpha 128)
• Hardware: NVIDIA H200 SXM
• Precision: BF16

Quantization

For deployment on smaller GPUs, use 4-bit quantization:

from transformers import AutoModelForCausalLM, BitsAndBytesConfig

model = AutoModelForCausalLM.from_pretrained(
    "Choukrijer/buildwellai-qwen3-14b",
    quantization_config=BitsAndBytesConfig(load_in_4bit=True),
    device_map="auto"
)

License

Apache 2.0 (same as base model)