README.md

metadata

language: en
tags:
  - reranker
  - RAG
  - multimodal
  - vision-language
  - Qwen
license: cc-by-4.0
pipeline_tag: visual-document-retrieval

DocReRank: Multi-Modal Reranker

This is the official model from the paper:

📄 DocReRank: Single-Page Hard Negative Query Generation for Training Multi-Modal RAG Rerankers

See Project Page for more information.

---

✅ Model Overview

• Base model: Qwen/Qwen2-VL-2B-Instruct
• Architecture: Vision-Language reranker
• Fine-tuning method: PEFT (LoRA)
• Training data: Generated by Single-Page Hard Negative Query Generation Pipeline.
• Purpose: Improves second-stage reranking for Retrieval-Augmented Generation (RAG) in multimodal scenarios.

---

✅ How to Use

This adapter requires the base Qwen2-VL model.

from transformers import AutoProcessor, Qwen2VLForConditionalGeneration
from peft import PeftModel
import torch
from PIL import Image

# ✅ Load base model
base_model = Qwen2VLForConditionalGeneration.from_pretrained(
    "Qwen/Qwen2-VL-2B-Instruct",
    torch_dtype=torch.bfloat16,
    device_map="cuda"
)

# ✅ Load DocReRank adapter
model = PeftModel.from_pretrained(base_model, "DocReRank/DocReRank-Reranker").eval()

# ✅ Load processor
processor = AutoProcessor.from_pretrained("Qwen/Qwen2-VL-2B-Instruct")
processor.image_processor.min_pixels = 200704
processor.image_processor.max_pixels = 589824

# ✅ Define query and images
query_text = "What are the performances of DocReRank model on resturants and Biomedical benchmarks?"
# query_text = "Is there ablations results of DocReRank model?"

### Paper pages to rank
image_paths = [ "DocReRank_paper_page_2.png", "DocReRank_paper_page_4.png", "DocReRank_paper_page_6.png", "DocReRank_paper_page_8.png"]

# ✅ Reranking prompt template
def compute_score(image_path, query_text):
    image = Image.open(image_path)
    prompt = f"Assert the relevance of the previous image document to the following query, answer True or False. The query is: {query_text}"
    messages = [{"role": "user", "content": [{"type": "image", "image": image}, {"type": "text", "text": prompt}]}]
    
    # Tokenize
    text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
    inputs = processor(text=text, images=image, return_tensors="pt").to(model.device, torch.bfloat16)
    
    # Compute logits
    with torch.no_grad():
        outputs = model(**inputs)
        logits = outputs.logits[:, -1, :]
        true_id = processor.tokenizer.convert_tokens_to_ids("True")
        false_id = processor.tokenizer.convert_tokens_to_ids("False")
        probs = torch.softmax(logits[:, [true_id, false_id]], dim=-1)
        relevance_score = probs[0, 0].item()  # Probability of "True"
    
    return relevance_score

# ✅ Compute scores for both images
scores = [(img, compute_score(img, query_text)) for img in image_paths]

# ✅ Print results
for img, score in scores:
    print(f"Image: {img} | Relevance Score: {score:.4f}")

Citation

If you use this dataset, please cite:

@article{wasserman2025docrerank,
  title={DocReRank: Single-Page Hard Negative Query Generation for Training Multi-Modal RAG Rerankers},
  author={Wasserman, Navve and Heinimann, Oliver and Golbari, Yuval and Zimbalist, Tal and Schwartz, Eli and Irani, Michal},
  journal={arXiv preprint arXiv:2505.22584},
  year={2025}
}