SauerkrautLM-ColQwen3-2b-v0.1

VAGO Solutions Logo

🥇 Best 128-dim Model in Medium (1-3B) Category | +1.01 over ColQwen2

SauerkrautLM-ColQwen3-2b-v0.1 achieves 90.24 NDCG@5 on ViDoRe v1, making it the #1 in the Medium (1-3B) category among 128-dim models - a significant +1.01 improvement over the baseline ColQwen2-v1.0.

ViDoRe v1 Benchmark - 128-dim Models

🎯 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

🏆 Key Achievements

Benchmark	Score	Rank (128-dim)
ViDoRe v1	90.24	#5
MTEB v1+v2	81.02	#6
ViDoRe v3	54.32	#5

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

Model	Params	Dim	ViDoRe v1	MTEB v1+v2	ViDoRe v3
SauerkrautLM-ColQwen3-2b-v0.1 ⭐	2.2B	128	90.24	81.02	54.32
colqwen2-v1.0	2.2B	128	89.23	79.74	44.18
SauerkrautLM-ColQwen3-1.7b-Turbo-v0.1	1.7B	128	88.89	77.94	48.76

#1 in Medium category on all three benchmarks!

Detailed Benchmark Results

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

Task	Score
ArxivQA	91.24
DocVQA	65.06
InfoVQA	93.14
ShiftProject	88.74
SyntheticDocQA-AI	99.63
SyntheticDocQA-Energy	96.91
SyntheticDocQA-Gov	96.08
SyntheticDocQA-Health	99.26
TabFQuAD	90.32
TATDQA	82.06
Average	90.24

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

ViDoRe v1 Tasks:

Task	Score
ArxivQA	91.24
DocVQA	65.06
InfoVQA	93.14
ShiftProject	88.74
SyntheticDocQA-AI	99.63
SyntheticDocQA-Energy	96.91
SyntheticDocQA-Gov	96.08
SyntheticDocQA-Health	99.26
TabFQuAD	90.32
TATDQA	82.06

ViDoRe v2 Tasks (Multilingual):

Task	Score
ViDoRe-v2-2BioMed	58.62
ViDoRe-v2-2Econ	54.64
ViDoRe-v2-2ESG-HL	68.13
ViDoRe-v2-2ESG	50.40
Combined Average	81.02

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

Task	Score
ViDoRe-v3-CS	73.70
ViDoRe-v3-Energy	61.21
ViDoRe-v3-FinanceEn	54.30
ViDoRe-v3-FinanceFr	40.18
ViDoRe-v3-HR	52.97
ViDoRe-v3-Industry	44.01
ViDoRe-v3-Pharma	60.64
ViDoRe-v3-Physics	47.57
Average	54.32

Improvement over Baseline

Metric	ColQwen3-2b	ColQwen2-v1.0	Improvement
ViDoRe v1	90.24	89.23	+1.01
MTEB v1+v2	81.02	79.74	+1.28
ViDoRe v3	54.32	44.18	+10.14

📋 Summary Tables

128-dim Models Comparison

128-dim Models Summary

Comparison vs High-dim Models

High-dim Comparison

✨ Key Features

🥇 #1 in Medium Category: Best 1-3B model among 128-dim models
📈 +1.01 over ColQwen2: Significant improvement over baseline
💾 Consumer GPU Ready: Only ~4.4GB VRAM
⚡ Compact Embeddings: 128-dimensional
🌍 Multilingual: 6 languages (EN, DE, FR, ES, IT, PT)

Model Details

Property	Value
Base Model	Qwen/Qwen3-VL-2B
Parameters	2.2B
Embedding Dimension	128
VRAM (bfloat16)	~4.4 GB
Max Context Length	262,144 tokens
License	Apache 2.0

Training

Hardware & Configuration

Setting	Value
GPUs	4x NVIDIA RTX 6000 Ada (48GB)
Effective Batch Size	256
Precision	bfloat16

Datasets

Dataset	Type	Description
vidore/colpali_train_set	Public	ColPali training data
openbmb/VisRAG-Ret-Train-In-domain-data	Public	Visual RAG training data
llamaindex/vdr-multilingual-train	Public	Multilingual document retrieval
VAGO Multilingual Dataset 1	In-house	Proprietary multilingual document-query pairs
VAGO Multilingual Dataset 2	In-house	Proprietary multilingual document-query pairs

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-ColQwen3-2b-v0.1",
    model_kwargs={"dtype": "bfloat16"},
)

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([[15.1758,  9.4717],
#         [ 4.1470, 14.3262]], 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-ColQwen3-2b-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)

📊 Additional Benchmark Visualizations

MTEB v1+v2 Benchmark (128-dim Models)

MTEB v1+v2 Benchmark - 128-dim Models

ViDoRe v3 Benchmark (128-dim Models)

ViDoRe v3 Benchmark - 128-dim Models

Our Models vs High-dim Models

ViDoRe v1 - Our Models vs High-dim

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}
}