Viraj-Sawad/vlm2vec-student-project

VLM2Vec Student Project: Hardness-Aware Visual RAG

Working backbone: Qwen/Qwen2-VL-2B-Instruct

Phi-3.5-Vision (VLM2Vec-Full checkpoint) has broken vision embedding loading. Qwen2-VL-2B works reliably.

⚠️ Critical Finding

Qwen2-VL requires <|image_pad|> token prepended to every text prompt that includes an image. Without it, the model ignores the image and all embeddings become identical.

# CORRECT (for Qwen2-VL)
prompt = f"<|image_pad|>{your_text}"
inputs = processor(text=[prompt], images=[image], ...)

# WRONG (image silently ignored)
inputs = processor(text=[your_text], images=[image], ...)  # DON'T DO THIS

Also: use AutoModel (not AutoModelForCausalLM or AutoModelForVision2Seq) for Qwen2-VL embedding extraction.

Files (run in this order)

Order	File	Purpose
1	smoke_test.py	Verify model loads and distinguishes semantics
2	train_domain.py	Fine-tune with hardness-weighted loss
3	ingest_corpus.py	Render PDFs → embed pages → ChromaDB
4	evaluate_models.py	Compare baseline vs fine-tuned metrics
5	demo_backend.py	FastAPI retrieval server
6	demo_frontend.py	Gradio side-by-side comparison UI

The ONLY Rule

Run smoke_test.py first. If it fails, nothing else matters.

Download

git clone https://huggingface.co/datasets/Viraj-Sawad/vlm2vec-student-project

Quick Start (Colab)

!pip install -q transformers accelerate bitsandbytes pillow requests qwen-vl-utils

import torch
from transformers import AutoProcessor, AutoModel
from PIL import Image
import requests

model_name = "Qwen/Qwen2-VL-2B-Instruct"
processor = AutoProcessor.from_pretrained(model_name, trust_remote_code=True)
model = AutoModel.from_pretrained(model_name, torch_dtype=torch.bfloat16,
                                  device_map="auto", trust_remote_code=True)

url = "http://images.cocodataset.org/val2017/000000039769.jpg"
image = Image.open(requests.get(url, stream=True).raw).convert("RGB")

def embed(text, img):
    prompt = f"<|image_pad|>{text}"  # ← CRITICAL
    inputs = processor(text=[prompt], images=[img], return_tensors="pt", padding=True)
    inputs = {k: v.to(model.device) if isinstance(v, torch.Tensor) else v
              for k, v in inputs.items()}
    with torch.no_grad():
        out = model(**inputs, output_hidden_states=True, return_dict=True)
        h = out.hidden_states[-1]
        emb = h[torch.arange(h.size(0)), inputs["attention_mask"].sum(1)-1, :]
        emb = torch.nn.functional.normalize(emb, p=2, dim=-1)
    return emb

emb_q = embed("Find images of cats sleeping", image)
emb_g = embed("Two cats sleeping on a pink couch", image)
emb_b = embed("A diagram of a neural network architecture", image)

print(f"Good: {torch.matmul(emb_q, emb_g.t()).item():.4f}")
print(f"Bad:  {torch.matmul(emb_q, emb_b.t()).item():.4f}")