eastrus-vl-qwen3-2b

A fine-tuned multimodal (vision-language) model based on Qwen3-VL-2B-Instruct, trained for cattle estrus-related vulval image assessment. The model produces symptom-by-symptom observations and a single confidence-style score, rather than a hard binary decision.

This repo contains the fully merged model weights in Hugging Face Transformers format (16-bit). For a quantized GGUF version for use with llama.cpp, see the companion repo.

Model details

• Base model: Qwen3-VL-2B-Instruct (VLM)
• Fine-tuning: Unsloth (LoRA, merged back into base weights)
• Format: Hugging Face Transformers (safetensors, 16-bit)

Intended use

• Primary: Assistive analysis of cattle vulval imagery for estrus-related visual signs (educational/research workflow support).
• Not a medical device: Outputs should not be used as the sole basis for veterinary diagnosis, treatment, or critical farm management decisions.

Output format

The model generates:

• A structured, human-readable symptom list (mucus color, swelling severity, redness severity, moisture level, mucus viscosity, tissue turgidity)
• A single confidence-style score (0-100%)
• A JSON structured summary of all observed symptoms

How to run (Transformers)

from transformers import AutoModelForImageTextToText, AutoProcessor
from PIL import Image

model = AutoModelForImageTextToText.from_pretrained("prapaa/eastrus-vl-qwen3-2b", torch_dtype="auto", device_map="auto")
processor = AutoProcessor.from_pretrained("prapaa/eastrus-vl-qwen3-2b")

image = Image.open("cow.jpg")
messages = [
    {"role": "system", "content": [{"type": "text", "text": "You are an expert veterinary AI..."}]},
    {"role": "user", "content": [
        {"type": "image", "image": image},
        {"type": "text", "text": "Analyze this cattle vulval image."},
    ]},
]
text = processor.apply_chat_template(messages, add_generation_prompt=True)
inputs = processor(images=[image], text=[text], return_tensors="pt").to(model.device)
output = model.generate(**inputs, max_new_tokens=384)
print(processor.decode(output[0], skip_special_tokens=True))