Construction Site Tagger — Thesis Model
A LoRA adapter for Qwen2-VL-7B-Instruct fine-tuned to tag construction-site photographs with infrastructure labels.
This is the thesis model — the version evaluated and reported in the research study.
The production successor is alexha11/construction-tagger-soupR.
Performance
| Eval set | Score |
|---|---|
| Thesis frozen set (original labels) | 0.866 |
| Thesis reported accuracy | 85.2 % |
Tags recognised
shallow_trench · medium_trench · deep_trench · cable_protection ·warning_tape · vegetated_ground · cable_drum · junction_box ·manhole · telecom_duct · electricity_duct
Adapter details
| Parameter | Value |
|---|---|
| Rank (r) | 16 |
| Alpha | 32 |
| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
| Training images | 1 155 |
Usage
from transformers import AutoProcessor, Qwen2VLForConditionalGeneration
from peft import PeftModel
base = Qwen2VLForConditionalGeneration.from_pretrained(
"Qwen/Qwen2-VL-7B-Instruct", torch_dtype="auto", device_map="auto"
)
model = PeftModel.from_pretrained(base, "alexha11/construction-tagger-thesis")
processor = AutoProcessor.from_pretrained("alexha11/construction-tagger-thesis")
Live demo
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support