SauerkrautLM-ColQwen3-1.7b-Turbo-v0.1

⚡ Turbo Edition | 23% Smaller, 88.89 ViDoRe v1

SauerkrautLM-ColQwen3-1.7b-Turbo-v0.1 is a pruned and optimized model created by applying structured pruning to Qwen3-VL-2B. Despite being 23% smaller, it achieves 88.89 NDCG@5 on ViDoRe v1 - still beating ColPali-v1.3 (84.75) by a large margin!

🎯 Why Visual Document Retrieval?

Traditional OCR-based retrieval loses layout, tables, and visual context. Our visual approach:

• ✅ No OCR errors - Direct visual understanding
• ✅ Layout-aware - Understands tables, forms, charts
• ✅ End-to-end - Single model, no pipeline complexity

✨ What Makes This "Turbo"?

🏆 Benchmark Results

Medium Category Comparison (1-3B, 128-dim)

vs. ColPali Baseline

Turbo model beats ColPali-v1.3 by +4.14 points with 42% fewer parameters!

Detailed Benchmark Results

📊 ViDoRe v1 (NDCG@5) - Click to expand

Task	Score
ArxivQA	91.09
DocVQA	60.98
InfoVQA	92.08
ShiftProject	87.76
SyntheticDocQA-AI	98.16
SyntheticDocQA-Energy	97.49
SyntheticDocQA-Gov	94.88
SyntheticDocQA-Health	98.26
TabFQuAD	87.03
TATDQA	81.19
Average	88.89

📊 MTEB v1+v2 (NDCG@5) - Click to expand

ViDoRe v1 Tasks:

Task	Score
ArxivQA	91.09
DocVQA	60.98
InfoVQA	92.08
ShiftProject	87.76
SyntheticDocQA-AI	98.16
SyntheticDocQA-Energy	97.49
SyntheticDocQA-Gov	94.88
SyntheticDocQA-Health	98.26
TabFQuAD	87.03
TATDQA	81.19

ViDoRe v2 Tasks (Multilingual):

Task	Score
ViDoRe-v2-2BioMed	53.92
ViDoRe-v2-2Econ	47.85
ViDoRe-v2-2ESG-HL	57.23
ViDoRe-v2-2ESG	43.27
Combined Average	77.94

📊 ViDoRe v3 (NDCG@10) - Click to expand

Task	Score
ViDoRe-v3-CS	67.14
ViDoRe-v3-Energy	56.48
ViDoRe-v3-FinanceEn	46.17
ViDoRe-v3-FinanceFr	33.73
ViDoRe-v3-HR	46.66
ViDoRe-v3-Industry	39.44
ViDoRe-v3-Pharma	55.31
ViDoRe-v3-Physics	45.14
Average	48.76

📋 Summary Tables

128-dim Models Comparison

Comparison vs High-dim Models

⚙️ Pruning Methodology

Structured Pruning

1. Layer Pruning: Removed less important transformer layers based on gradient-based importance scoring
2. Intermediate Size Reduction: Reduced FFN intermediate dimensions
3. Result: 23% parameter reduction (2.2B → 1.7B)

Recovery Training with mMARCO

After pruning, the model underwent recovery training:

Pruned Model → mMARCO Pre-training → Fine-tuning → Final Model

The mMARCO pre-training was crucial to "heal" the model after pruning.

Model Details

Training

Hardware & Configuration

Training Pipeline

1. Phase 1: Structured Pruning (gradient-based importance)
2. Phase 2: mMARCO Recovery Training
3. Phase 3: Retrieval Fine-tuning on standard datasets

Installation & Usage

Sentence Transformers

This model can be used with Sentence Transformers as a multi-vector (ColBERT-style late interaction) retriever via the MultiVectorEncoder:

pip install "sentence-transformers[image]>=6.0.0"

from sentence_transformers import MultiVectorEncoder

model = MultiVectorEncoder("VAGOsolutions/SauerkrautLM-ColQwen3-1.7b-Turbo-v0.1")

queries = [
    "What is the variable represented on the y-axis of the graph?",
    "Total outlay is maximum in which year?",
]
images = [
    "https://huggingface.co/datasets/sentence-transformers/example-documents/resolve/main/doc1.jpg",
    "https://huggingface.co/datasets/sentence-transformers/example-documents/resolve/main/doc2.jpg",
]

query_embeddings = model.encode_query(queries)
image_embeddings = model.encode_document(images)
print(query_embeddings[0].shape, image_embeddings[0].shape)
# torch.Size([25, 128]) torch.Size([1251, 128])

# Diagonal should have higher scores
scores = model.similarity(query_embeddings, image_embeddings)
print(scores)
# tensor([[14.5811, 11.6055],
#         [ 7.1665, 14.7393]], device='cuda:0')

SauerkrautLM ColPali

⚠️ Important: Install our package first before loading the model:

pip install git+https://github.com/VAGOsolutions/sauerkrautlm-colpali

import torch
from PIL import Image
from sauerkrautlm_colpali.models import ColQwen3, ColQwen3Processor

model_name = "VAGOsolutions/SauerkrautLM-ColQwen3-1.7b-Turbo-v0.1"

model = ColQwen3.from_pretrained(
    model_name,
    torch_dtype=torch.bfloat16,
    attn_implementation="flash_attention_2",
    device_map="cuda:0",
).eval()

processor = ColQwen3Processor.from_pretrained(model_name)

images = [Image.open("document.png")]
queries = ["What is the main topic?"]

batch_images = processor.process_images(images).to(model.device)
batch_queries = processor.process_queries(queries).to(model.device)

with torch.no_grad():
    image_embeddings = model(**batch_images)
    query_embeddings = model(**batch_queries)

scores = processor.score(query_embeddings, image_embeddings)

When to Use Turbo

✅ Choose Turbo when:

• Running on limited GPU memory (< 4GB available)
• Need faster inference
• Deploying on edge devices
• Cost optimization is priority

❌ Choose 2B instead when:

• Maximum accuracy required
• Memory is not a constraint

📊 Additional Benchmark Visualizations

MTEB v1+v2 Benchmark (128-dim Models)

ViDoRe v3 Benchmark (128-dim Models)

Our Models vs High-dim Models

Citation

@misc{sauerkrautlm-colpali-2025,
  title={SauerkrautLM-ColPali: Multi-Vector Vision Retrieval Models},
  author={David Golchinfar},
  organization={VAGO Solutions},
  year={2025},
  url={https://github.com/VAGOsolutions/sauerkrautlm-colpali}
}

Contact

• VAGO Solutions: https://vago-solutions.ai
• GitHub: https://github.com/VAGOsolutions