mcdse-2b-v1
mcdse-2b-v1 is an experimental model designed for multilingual visual document retrieval.
This model allows you to embed page/slide screenshots and query them using natural language. Whether it's tables, graphs, charts, schemas, images, or text, mcdse-2b-v1 encodes everything into a single embedding vector, eliminating the need for traditional OCR, document layout analysis, reading order detection, chunking, table/formula extraction...
Understands ๐ฎ๐น Italian, ๐ช๐ธ Spanish, ๐ฌ๐ง English, ๐ซ๐ท French and ๐ฉ๐ช German
Matryoshka Representation Learning: shrink embeddings from 1536 to 256 dimensions while maintaining 95% of the quality. A 6x reduction with negligible impact on performance!
Top-tier Binarization: 768-dimensional binary vectors retain 99% retrieval quality of the original 1536-dimensional float vectors. With binary vectors, you can encode 100 million multilingual pages in just 10GB.
Fast vLLM inference: run inference on vLLM and efficiently serve embeddings at scale, production ready.
For more information about this model or how it was trained, visit the announcement blogpost.
Usage
Initialize model and processor
from transformers import AutoProcessor, Qwen2VLForConditionalGeneration
from PIL import Image
import torch
import math
model = Qwen2VLForConditionalGeneration.from_pretrained(
'marco/mcdse-2b-v1',
attn_implementation="flash_attention_2",
torch_dtype=torch.bfloat16,
device_map="cuda:0"
).eval()
min_pixels = 1 * 28 * 28
max_pixels = 960 * 28 * 28
processor = AutoProcessor.from_pretrained(
'marco/mcdse-2b-v1',
min_pixels=min_pixels,
max_pixels=max_pixels
)
model.padding_side = "left"
processor.tokenizer.padding_side = "left"
document_prompt = "<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n<|im_start|>user\n<|vision_start|><|image_pad|><|vision_end|>What is shown in this image?<|im_end|>\n<|endoftext|>"
query_prompt = "<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n<|im_start|>user\n<|vision_start|><|image_pad|><|vision_end|>Query: %s<|im_end|>\n<|endoftext|>"
Encode queries
def encode_queries(queries: list[str], dimension: int):
dummy_image = Image.new('RGB', (56, 56))
inputs = processor(
text=[query_prompt % x for x in queries],
images=[dummy_image for _ in queries],
videos=None,
padding='longest',
return_tensors='pt'
).to('cuda:0')
cache_position = torch.arange(0, len(queries))
inputs = model.prepare_inputs_for_generation(
**inputs, cache_position=cache_position, use_cache=False)
with torch.no_grad():
output = self.model(
**inputs,
return_dict=True,
output_hidden_states=True
)
embeddings = output.hidden_states[-1][:, -1]
return torch.nn.functional.normalize(embeddings[:, :dimension], p=2, dim=-1)
Encode documents
def round_by_factor(number: float, factor: int) -> int:
return round(number / factor) * factor
def ceil_by_factor(number: float, factor: int) -> int:
return math.ceil(number / factor) * factor
def floor_by_factor(number: float, factor: int) -> int:
return math.floor(number / factor) * factor
def smart_resize(height: int, width: int) -> tuple[int, int]:
h_bar = max(28, round_by_factor(height, 28))
w_bar = max(28, round_by_factor(width, 28))
if h_bar * w_bar > max_pixels:
beta = math.sqrt((height * width) / max_pixels)
h_bar = floor_by_factor(height / beta, 28)
w_bar = floor_by_factor(width / beta, 28)
elif h_bar * w_bar < min_pixels:
beta = math.sqrt(min_pixels / (height * width))
h_bar = ceil_by_factor(height * beta, 28)
w_bar = ceil_by_factor(width * beta, 28)
return h_bar, w_bar
def resize(image: Image.Image):
new_size = smart_resize(image.height, image.width)
return image.resize(new_size)
def encode_documents(documents: list[Image.Image], dimension: int):
inputs = processor(
text=[document_prompt] * len(documents),
images=[resize(x) for x in documents],
videos=None,
padding='longest',
return_tensors='pt'
).to('cuda:0')
cache_position = torch.arange(0, len(queries))
inputs = model.prepare_inputs_for_generation(
**inputs, cache_position=cache_position, use_cache=False)
with torch.no_grad():
output = self.model(
**inputs,
return_dict=True,
output_hidden_states=True
)
embeddings = output.hidden_states[-1][:, -1]
return torch.nn.functional.normalize(embeddings[:, :dimension], p=2, dim=-1)
vLLM
This model supports vLLM, visit the announcement blogpost to know more.
Results
Given the scarcity of publicly available datasets for multilingual document image retrieval, the model has been evaluated using a custom-built dataset. This eval dataset was specifically designed to benchmark the model's performance across various languages.
NDCG@5 (float)
Average | English | Italian | Spanish | French | German | |
---|---|---|---|---|---|---|
1536 dimensions | ||||||
dse-qwen2-2b-mrl-v1 | 79.5 | 79.2 | 80.2 | 77.9 | 80.6 | 79.6 |
mcdse-2b-v1 | 82.2 | 80.8 | 81.2 | 80.7 | 84.5 | 83.8 |
+3.28% | +1.98% | +1.23% | +3.47% | +4.62% | +5.01% | |
1024 dimensions | ||||||
dse-qwen2-2b-mrl-v1 | 78.3 | 78.8 | 78.5 | 76.5 | 80 | 77.5 |
mcdse-2b-v1 | 81.7 | 80 | 80.2 | 80.1 | 84 | 84.3 |
+4.23% | +1.75% | +2.12% | +4.49% | +4.76% | +8.07% | |
768 dimensions | ||||||
dse-qwen2-2b-mrl-v1 | 77.8 | 78.4 | 78.3 | 75.6 | 80.8 | 75.9 |
mcdse-2b-v1 | 81.1 | 79.6 | 79.9 | 79.2 | 83.3 | 83.3 |
+4.02% | +1.51% | +2.00% | +4.55% | +3.00% | +8.88% | |
512 dimensions | ||||||
dse-qwen2-2b-mrl-v1 | 76.2 | 77.6 | 75.9 | 73.1 | 79.2 | 75.2 |
mcdse-2b-v1 | 79.3 | 78.5 | 79.1 | 75.8 | 81.4 | 81.7 |
+3.91% | +1.15% | +4.05% | +3.56% | +2.70% | +7.96% | |
384 dimensions | ||||||
dse-qwen2-2b-mrl-v1 | 75.7 | 76.2 | 75.5 | 74.6 | 78.4 | 74 |
mcdse-2b-v1 | 78.8 | 77.5 | 78.5 | 76.1 | 80.4 | 81.4 |
+3.86% | +1.68% | +3.82% | +1.97% | +2.49% | +9.09% | |
256 dimensions | ||||||
dse-qwen2-2b-mrl-v1 | 73.5 | 74.5 | 73.6 | 70.6 | 74.8 | 73.8 |
mcdse-2b-v1 | 78.1 | 78.5 | 77.6 | 76.2 | 80.1 | 77.9 |
+5.89% | +5.10% | +5.15% | +7.35% | +6.62% | +5.26% |
NDCG@5 (binary)
Average | English | Italian | Spanish | French | German | |
---|---|---|---|---|---|---|
1536 dimensions | ||||||
dse-qwen2-2b-mrl-v1 | 75.0 | 75.8 | 75.4 | 72.4 | 78.1 | 73.2 |
mcdse-2b-v1 | 80.6 | 79.5 | 76.9 | 81.9 | 83.7 | 80.8 |
+6.93% | +4.65% | +1.95% | +11.60% | +6.69% | +9.41% | |
1024 dimensions | ||||||
dse-qwen2-2b-mrl-v1 | 72.2 | 74.8 | 71 | 70.8 | 74.6 | 69.6 |
mcdse-2b-v1 | 79.3 | 78.4 | 75.4 | 80.8 | 82.6 | 79.5 |
+9.05% | +4.59% | +5.84% | +12.38% | +9.69% | +12.45% | |
768 dimensions | ||||||
dse-qwen2-2b-mrl-v1 | 70.1 | 71.7 | 69.3 | 69.8 | 73.7 | 65.9 |
mcdse-2b-v1 | 78.8 | 77.1 | 75.4 | 80 | 83 | 78.5 |
+11.07% | +7.00% | +8.09% | +12.75% | +11.20% | +16.05% | |
512 dimensions | ||||||
dse-qwen2-2b-mrl-v1 | 66.5 | 70 | 65.4 | 63.7 | 70.2 | 63 |
mcdse-2b-v1 | 76.6 | 74.8 | 74.2 | 77.7 | 80.9 | 75.3 |
+13.21% | +6.42% | +11.86% | +18.02% | +13.23% | +16.33% | |
384 dimensions | ||||||
dse-qwen2-2b-mrl-v1 | 61.1 | 62.7 | 58.5 | 58.6 | 65.1 | 60.8 |
mcdse-2b-v1 | 74.3 | 74.5 | 71.4 | 77.2 | 75.2 | 73 |
+17.67% | +15.84% | +18.07% | +24.09% | +13.43% | +16.71% | |
256 dimensions | ||||||
dse-qwen2-2b-mrl-v1 | 54.3 | 59 | 56.5 | 53.6 | 53 | 49.6 |
mcdse-2b-v1 | 70.9 | 72.6 | 66.4 | 73.5 | 72.6 | 69.2 |
+23.31% | +18.73% | +14.91% | +27.07% | +27.00% | +28.32% |
- Downloads last month
- 369