Jsohal174/hawkeye-warehouse-safety-lora

Warehouse Safety Hazard Detection, LoRA Adapter

LoRA adapter for warehouse safety hazard classification, trained on synthetic warehouse data on top of unsloth/Qwen3.5-2B.

If you just want to run the model, use the Q4 GGUF release instead. This repo is the raw adapter for people who want to merge it themselves or continue training.

Research project at the University of Guelph supervised by Dr. John Akinyemi.

Adapter configuration

Setting	Value
Base model	unsloth/Qwen3.5-2B
Rank (r)	16
Alpha	32
Dropout	0.05
Bias	none
Task type	CAUSAL_LM
Targets	attention and MLP projections, vision and language towers
Adapter size	~88 MB

Usage

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoProcessor

base = AutoModelForCausalLM.from_pretrained("unsloth/Qwen3.5-2B", device_map="auto")
model = PeftModel.from_pretrained(base, "Jsohal174/hawkeye-warehouse-safety-lora")
processor = AutoProcessor.from_pretrained("unsloth/Qwen3.5-2B")

Categories

spill, forklift_violation, improper_stacking, obstacle, safe

Results

Held-out test set, n=186 images. Measured after merging and quantizing to Q4_K_M, running on a Mac Mini M4.

Configuration	Accuracy	Latency
Trained, Q4 no-think (deployed)	91.9%	~3 s
Trained, Q4 think	92.5%	~4.2 s
Base, zero-shot	65.3%	~3.1 s

Training delta: +26.6 percentage points over the same base, same prompt, same hardware.

Training

2,521 conversations from 727 images, 3.5 hours on a free Kaggle T4.

Related

• Q4 GGUF, Ollama-ready: https://huggingface.co/Jsohal174/hawkeye-warehouse-safety-gguf
• Full precision merged: https://huggingface.co/Jsohal174/hawkeye-warehouse-safety-merged
• Dataset, 913 labeled images: https://huggingface.co/datasets/Jsohal174/warehouse-safety-hazard-dataset
• Code and benchmarks: https://github.com/Jsohal174/safety-hazard-detection

Citation

@misc{sohal2026synthetic,
  author = {Sohal, Jaskirat Singh},
  title  = {Synthetic Specialization: Training and Compressing Foundation Models for Domain-Specific Tasks},
  year   = {2026},
  howpublished = {University of Guelph},
  note   = {Supervisor: Dr. John Akinyemi},
  url    = {https://github.com/Jsohal174/safety-hazard-detection}
}