Qwen2.5-1.5B BuildEng V7
Repository: Irfanuruchi/qwen2.5-1.5b-buildeng-v7-lora
This model is a building-engineering focused LoRA fine-tune of Qwen2.5-1.5B-Instruct.
V7 was built to move away from only direct engineering calculations and push the model more toward engineering reasoning behavior. Earlier versions could solve formulas, but still sometimes accepted wrong assumptions, bad unit conversions, unrealistic HVAC values, or incomplete design prompts.
The main goal of V7 was making the model more careful instead of only making it sound technical.
What I Used
Base model:
Qwen/Qwen2.5-1.5B-Instruct
Training method:
LoRA fine-tuning with Unsloth
Dataset used:
buildeng-v7-15k
The dataset was generated using custom Python generators focused on engineering calculations, sanity checks, formula correction, and refusal behavior.
Main Focus Of V7
V7 focuses heavily on HVAC airflow reasoning, unit consistency, m³/h to m³/s conversion, duct sizing logic, formula correction, engineering sanity checks, ventilation reasoning, and missing-input refusal behavior.
One of the biggest goals during V7 development was stopping the model from sounding confident while still being wrong.
Earlier versions could still:
accept wrong duct equations
calculate ACH airflow from floor area only
accept unrealistic HVAC velocities
guess missing engineering values
confuse airflow units
V7 adds much more correction-focused and sanity-focused training around those exact problems.
Example Behavior
Example:
User:
7200 m³/h equals 7200 m³/s, right?
Assistant:
No. 7200 m³/h ÷ 3600 = 2 m³/s.
Another example:
User:
Calculate duct diameter.
Assistant:
More information is required before sizing the duct. Airflow and target velocity are needed.
Engineering Areas
The model was trained mainly around simplified building-engineering reasoning tasks including HVAC airflow calculations, duct sizing, ventilation logic, pressure reasoning, hydronic sanity checks, footing pressure comparison, engineering unit conversion, formula correction, and missing-input engineering prompts.
Limitations
This model is not a replacement for professional engineering software or real engineering review.
The training data is synthetic and simplified. Full ASHRAE, Eurocode, or professional design standards are not fully implemented.
Outputs should always be checked by a qualified engineer.
The model can still make arithmetic mistakes or produce incorrect engineering assumptions in some cases, especially on more complex problems.
Future Improvements
V7 is an intermediate version before larger datasets and stronger engineering-specialized models.
Future work includes larger datasets, stronger arithmetic consistency, more structural reasoning, richer HVAC workflows, multi-turn engineering reasoning, and larger engineering-specialized models including 32B systems.
Training Notes
One important part during V7 development was using failed model outputs as feedback for future dataset generation.
When the model repeatedly failed certain engineering cases, additional repair-focused synthetic samples were added around those exact weak areas.
This made the tuning process much more iterative instead of only generating random engineering calculations.
Usage
This repository contains a LoRA adapter, not full standalone model weights.
Base model required:
Qwen/Qwen2.5-1.5B-Instruct
Author
Irfan Uruchi
Part of my ongoing work on domain-specialized engineering language models and lightweight engineering reasoning systems.