Qwen2.5-1.5B Building Engineering Precheck LoRA (V5)

Repository: Irfanuruchi/qwen2.5-1.5b-buildeng-precheck-lora-v5
Base model: Qwen2.5-1.5B-Instruct
Fine-tuning method: LoRA (Unsloth)


Overview

This repository provides a LoRA adapter for Qwen2.5-1.5B-Instruct fine-tuned for building engineering reasoning tasks.

The goal of this project is to create a lightweight domain-specific assistant capable of performing engineering pre-check calculations and sanity validation.

The model focuses on early-stage engineering estimation, supporting tasks such as:

• HVAC duct sizing
• Hydronic pipe sizing
• Fan and pump power estimation
• Heat transfer calculations
• Structural pre-check calculations
• Engineering unit sanity checks
• Detection of missing or unrealistic design inputs

This model acts as an engineering assistant for early concept validation, not as a replacement for professional engineering verification.


Training Details

Base Model: Qwen2.5-1.5B-Instruct
Framework: Unsloth
Fine-tuning Method: LoRA

Training Configuration

Parameter Value
LoRA rank 32
LoRA alpha 32
Quantization 4-bit
Sequence length 1536
Optimizer AdamW 8-bit
Learning rate 1.5e-4
Scheduler Cosine
Training steps 3000
Train dataset ~57k samples
Validation dataset ~3k samples

The dataset consists of synthetic engineering problems generated using physics-based equations covering HVAC, hydronics, heat transfer, and structural estimation.


Example Capabilities

HVAC duct sizing

Input

Airflow = 3500 m³/h
Velocity = 6 m/s

Output

Select round duct ≈ Ø480 mm

Steps

Q = 0.972 m³/s
A = 0.162 m²
d ≈ 480 mm


Pump power estimation

Input

Flow = 4 m³/h
Head = 20 m
Efficiency = 0.7

Output

Pump power ≈ 0.31 kW

Steps

P = ρ g Q H / η


Structural stress calculation

Input

Moment = 20 kNm
Section modulus = 250 cm³

Output

Bending stress ≈ 80 MPa


Engineering sanity check

Input

Size round duct
Airflow = 0.3 m³/h
Velocity = 5 m/s

Output

Airflow extremely small for HVAC systems.
Likely unit mismatch — confirm m³/h vs m³/s.


Engineering Tasks Covered

HVAC

• duct sizing
• fan power estimation
• cooling coil loads

Hydronics

• pipe velocity sizing
• pump power estimation
• water flow calculations

Heat Transfer

• wall heat loss estimation

Structural Pre-checks

• beam bending stress
• beam deflection estimation
• tributary load calculation
• slab load estimation

Engineering Safety Checks

• unrealistic velocities
• unrealistic ΔT values
• unit mismatches
• missing critical inputs


Limitations

This model is experimental and intended for educational or exploratory use.

Known limitations include:

• calculations may require verification
• reasoning depth is limited by model size
• structural calculations are simplified
• model should not be used for final engineering design

All results must be validated using professional engineering standards (ASHRAE, Eurocode, etc.).


Usage

Load the base model and apply the LoRA adapter.

Example using Unsloth:

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="Qwen/Qwen2.5-1.5B-Instruct",
    load_in_4bit=True
)

model.load_adapter("path_to_adapter")

FastLanguageModel.for_inference(model)

Training Dataset

The model was trained on the following dataset:

https://huggingface.co/datasets/Irfanuruchi/building-engineering-synthetic-dataset-v5


Intended Use

This model is designed for:

• educational engineering demonstrations
• quick engineering pre-check calculations
• concept-stage HVAC sizing
• structural estimation checks

It should never be used for safety-critical engineering decisions without verification.


Future Work

Planned improvements include:

• expanding dataset beyond 100k engineering problems
• deeper structural engineering reasoning
• multi-step engineering calculation chains
• integration of engineering standards (ASHRAE, Eurocode)
• improved unit consistency detection


Version History

V1 – Initial experiments
V2 – Dataset structure improvements
V3 – HVAC reasoning tasks
V4 – Engineering pre-check assistant (5k dataset)
V5 – Expanded dataset (
60k), HVAC + structural reasoning + engineering sanity checks


Author

Irfan Uruchi

This project explores lightweight domain-specialized LLMs for engineering reasoning and pre-design validation (also used as testing for my 32B parameter LLM launching by June).

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support