mmrech/pitvqa-qwen2vl-merged

PitVQA Merged Model

A ready-to-use version of PitVQA with the Stage 4 (Unified) adapter merged into the base weights. No adapter loading required - just load and run.

Why Use This Model?

Feature	This (Merged)	Adapter Version
Setup complexity	Simple	Requires PEFT
Load time	Faster	Slower
Flexibility	Single task mode	Switch adapters
Best for	Production deployment	Research/experimentation

Quick Start

from transformers import Qwen2VLForConditionalGeneration, AutoProcessor
import torch

# Load merged model directly - no adapter loading needed!
model = Qwen2VLForConditionalGeneration.from_pretrained(
    "mmrech/pitvqa-qwen2vl-merged",
    torch_dtype=torch.bfloat16,
    device_map="auto"
)
processor = AutoProcessor.from_pretrained("mmrech/pitvqa-qwen2vl-merged")

# Run inference
from PIL import Image
image = Image.open("surgical_frame.jpg")

messages = [{"role": "user", "content": [
    {"type": "image", "image": image},
    {"type": "text", "text": "Point to the suction device in this surgical image."}
]}]

text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = processor(text=[text], images=[image], return_tensors="pt").to(model.device)

output = model.generate(**inputs, max_new_tokens=128)
print(processor.decode(output[0], skip_special_tokens=True))
# Output: <point x='75.8' y='75.1'>suction device</point>

With 4-bit Quantization (Lower Memory)

from transformers import BitsAndBytesConfig

bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16,
)

model = Qwen2VLForConditionalGeneration.from_pretrained(
    "mmrech/pitvqa-qwen2vl-merged",
    quantization_config=bnb_config,
    device_map="auto"
)

Supported Tasks

This merged model supports ALL tasks from Stage 4 training:

Point Localization

Prompt: "Point to the suction device in this surgical image."
Output: <point x='75.8' y='75.1'>suction device</point>

Bounding Box Detection

Prompt: "Draw a bounding box around the tumor region."
Output: <box x1='30' y1='30' x2='70' y2='70'>tumor region</box>

Phase Classification

Prompt: "What surgical phase is shown?"
Output: sellar phase

Supported phases: nasal, sellar, tumor_removal, closure

Free-form Queries

Prompt: "What instruments are visible in this surgical scene?"
Output: The image shows a suction device positioned in the surgical field...

Model Details

• Source: Merged from mmrech/pitvqa-qwen2vl-unified-v2 Stage 4 adapter
• Base: Qwen/Qwen2-VL-2B-Instruct
• Parameters: ~2B (full model, adapter merged)
• Training: SFT with LoRA, then merged with merge_and_unload()

Related Resources

• Adapter Version: pitvqa-qwen2vl-unified-v2 (for multi-stage switching)
• Dataset: pitvqa-comprehensive-spatial
• Demo: HuggingFace Space

Citation

@misc{pitvqa2026,
  title={PitVQA: Multi-Task Vision-Language Model for Pituitary Surgery},
  author={Matheus Rech},
  year={2026},
  url={https://huggingface.co/mmrech/pitvqa-qwen2vl-merged}
}

License

Apache 2.0