WueNLP/centurio_qwen

Centurio Qwen

Model Details

Model Description

• Model type: Centurio is an open-source multilingual large vision-language model.
• Training Data: COMING SOON
• Languages: The model was trained with the following 100 languages: af, am, ar, ar-eg, as, azb, be, bg, bm, bn, bo, bs, ca, ceb, cs, cy, da, de, du, el, en, eo, es, et, eu, fa, fi, fr, ga, gd, gl, ha, hi, hr, ht, hu, id, ig, is, it, iw, ja, jv, ka, ki, kk, km, ko, la, lb, ln, lo, lt, lv, mi, mr, ms, mt, my, no, oc, pa, pl, pt, qu, ro, ru, sa, sc, sd, sg, sk, sl, sm, so, sq, sr, ss, sv, sw, ta, te, th, ti, tl, tn, tpi, tr, ts, tw, uk, ur, uz, vi, war, wo, xh, yo, zh, zu
• License: This work is released under the Apache 2.0 license.

Model Sources

• Repository: gregor-ge.github.io/Centurio
• Paper: arXiv

Uses

Direct Use

The model can be used directly through the transformers library with our custom code.

from transformers import AutoModelForCausalLM, AutoProcessor
import timm
from PIL import Image    
import requests

url = "https://upload.wikimedia.org/wikipedia/commons/b/bd/Golden_Retriever_Dukedestiny01_drvd.jpg"
image = Image.open(requests.get(url, stream=True).raw)

model_name = "WueNLP/centurio_qwen"

processor = AutoProcessor.from_pretrained(model_name, trust_remote_code=True)

## Appearance of images in the prompt are indicates with '<image_placeholder>'!
prompt = "<image_placeholder>\nBriefly describe the image in German."

messages = [
    {"role": "system", "content": "You are a helpful assistant."},  # This is the system prompt used during our training.
    {"role": "user", "content": prompt}
]

text = processor.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)

model = AutoModelForCausalLM.from_pretrained(
    model_name,
    trust_remote_code=True
)

model_inputs = processor(text=[text], images=[image] return_tensors="pt").to(model.device)

generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=128
)

generated_ids = [
    output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
]

response = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]

Multiple Images

We natively support multi-image inputs. You only have to 1) include more <image_placeholder> while 2) passing all images of the entire batch as a flat list:

[...]
# Variables reused from above.

processor.tokenizer.padding_side = "left" # default is 'right' but has to be 'left' for batched generation to work correctly!

image_multi_1, image_multi_2 = [...] # prepare additional images

prompt_multi = "What is the difference between the following images?\n<image_placeholder><image_placeholder>\nAnswer in German."

messages_multi = [
    {"role": "system", "content": "You are a helpful assistant."}, 
    {"role": "user", "content": prompt_multi}
]

text_multi = processor.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)

model_inputs = processor(text=[text, text_multi], images=[image, image_multi_1, image_multi_2] return_tensors="pt").to(model.device)

generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=128
)

[...]

Bias, Risks, and Limitations

• General biases, risks, and limitations of large vision-language models like hallucinations or biases from training data apply.
• This is a research project and not recommended for production use.
• Multilingual: Performance and generation quality can differ widely between languages.
• OCR: Model struggles both with small text and writing in non-Latin scripts.

Citation

BibTeX:

@article{centurio2025,
  author       = {Gregor Geigle and
                  Florian Schneider and
                  Carolin Holtermann and
                  Chris Biemann and
                  Radu Timofte and
                  Anne Lauscher and
                  Goran Glava\v{s}},
  title        = {Centurio: On Drivers of Multilingual Ability of Large Vision-Language Model},
  journal      = {arXiv},
  volume       = {abs/2501.05122},
  year         = {2025},
  url          = {https://arxiv.org/abs/2501.05122},
  eprinttype    = {arXiv},
  eprint       = {2501.05122},
}