Modelpublic

vorenthiclabs/vorenthos-vision-coding-qwen25vl-7b

sourceHugging Faceapache-2.0updated 2mo agoView on Hugging Face

0likes7downloads

Model Card

Vision + Coding Model

Fine-tuned [unsloth/Qwen2-VL-7B-Instruct-bnb-4bit](https://huggingface.co/unsloth/Qwen2-VL-7B-Instruct-bnb-4bit) with Unsloth + QLoRA.

Qwen2.5-VL-7B is the top open-source VLM for code, OCR, and document understanding at the 7B scale. Fine-tuned here to read code screenshots, diagrams, and technical images and reproduce or explain the code they contain.

Training details

Setting	Value
Base model	`unsloth/Qwen2-VL-7B-Instruct-bnb-4bit`
Method	QLoRA (r=16, alpha=16)
Precision	float16 (merged)
Max seq length	2048
Steps	60
Date	2026-07-26

Usage

python

from transformers import Qwen2_5_VLForConditionalGeneration, AutoProcessor
from qwen_vl_utils import process_vision_info
import torch
 
model_id = "vorenthiclabs/vorenthos-vision-coding-qwen25vl-7b"
processor = AutoProcessor.from_pretrained(model_id)
model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
    model_id, torch_dtype=torch.float16, device_map="auto"
)
 
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "image": "https://your-code-screenshot.png"},
            {"type": "text",  "text": "Reproduce the code shown in this screenshot."}
        ]
    }
]
text   = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = processor(text=[text], images=[...], return_tensors="pt").to(model.device)
output = model.generate(**inputs, max_new_tokens=1024, temperature=0.1)
print(processor.decode(output[0], skip_special_tokens=True))

Best use cases

—Screenshot → working code
—UI mockup → HTML/CSS
—Math diagram → LaTeX / pseudocode
—Technical document → structured extraction