Model Card for RoGemma3-4B-Instruct

RoGemma3-4B-Instruct is a Romanian-adapted vision-language model built on top of google/gemma-3-4b-it. It was produced by continued supervised instruction tuning of the base Gemma 3 checkpoint on a Romanian multimodal SFT mixture covering general instruction following (LLaVA mix), captioning (Pixmo-Cap, Flickr30k-Cap), visual question answering (Pixmo-AA, Pixmo-Cap-QA, Flickr30k-QA), document and chart understanding (CoSyn, FinePDFs), and visual grounding (Pixmo-Points, Pixmo-Count). The model is intended for research on Romanian VLM capabilities.

Model Details

Model Description

• Developed by: OpenLLM-Ro
• Language(s): Romanian
• License: cc-by-nc-4.0
• Finetuned from model: google/gemma-3-4b-it
• Trained using:
  • OpenLLM-Ro/ro_sft_laion
  • OpenLLM-Ro/ro_sft_pixmo_cap
  • OpenLLM-Ro/ro_sft_flickr30k_cap
  • OpenLLM-Ro/ro_sft_llava_mix
  • OpenLLM-Ro/ro_sft_pixmo_aa
  • OpenLLM-Ro/ro_sft_pixmo_cap_qa
  • OpenLLM-Ro/ro_sft_flickr30k_qa
  • OpenLLM-Ro/ro_sft_cosyn
  • OpenLLM-Ro/ro_sft_finepdfs
  • OpenLLM-Ro/ro_sft_pixmo_points
  • OpenLLM-Ro/ro_sft_pixmo_count

Model Sources

• Repository: https://github.com/OpenLLM-Ro/LLaMA-Factory
• Paper: https://arxiv.org/abs/2605.31401

Intended Use

Intended Use Cases

RoGemma3-4B-Instruct is intended for research use on Romanian vision-language tasks — captioning, visual question answering, cultural understanding, OCR / document understanding, and visual grounding — and as a starting point for further Romanian VLM adaptation.

Out-of-Scope Use

Use in any manner that violates applicable laws or regulations (including trade-compliance laws), the project's license, or use in languages other than Romanian.

How to Get Started with the Model

import torch
from PIL import Image
from transformers import AutoProcessor, Gemma3ForConditionalGeneration

model = Gemma3ForConditionalGeneration.from_pretrained(
    "OpenLLM-Ro/RoGemma3-4B-Instruct",
    torch_dtype=torch.bfloat16,
    device_map="auto",
).eval()
processor = AutoProcessor.from_pretrained("OpenLLM-Ro/RoGemma3-4B-Instruct")

image = Image.open("example.jpg").convert("RGB")
question = "Descrie imaginea în detaliu."

messages = [
    {"role": "user", "content": [
        {"type": "image", "image": image},
        {"type": "text", "text": question},
    ]},
]
inputs = processor.apply_chat_template(
    messages,
    add_generation_prompt=True,
    tokenize=True,
    return_dict=True,
    return_tensors="pt",
).to(model.device, dtype=torch.bfloat16)

with torch.inference_mode():
    outputs = model.generate(**inputs, max_new_tokens=256, do_sample=False)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))

Benchmarks

All benchmarks below are evaluated in Romanian. Per-benchmark winners are shown in bold. Micro is the mean over individual benchmarks; Macro is the mean over capability groups.

Aggregate

General Understanding

Knowledge & Reasoning

Cultural

Generation & Open-ended

OCR & Documents

Grounding

Citation

@misc{masala2026intelegi,
      title={``\^{I}n\c{t}elegi Rom\^{a}ne\c{s}te?'' A Recipe for Romanian Vision-Language Models},
      author={Mihai Masala and Marius Leordeanu and Mihai Dascalu and Traian Rebedea},
      year={2026},
      eprint={2605.31401},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2605.31401},
}

@inproceedings{masala-etal-2024-vorbesti,
    title = "``Vorbeşti Româneşte?'' A Recipe to Train Powerful {R}omanian {LLM}s with {E}nglish Instructions",
    author = "Masala, Mihai and Ilie-Ablachim, Denis and Dima, Alexandru and Corlatescu, Dragos and Zavelca, Miruna and Olaru, Ovio and Terian, Simina and Terian, Andrei and Leordeanu, Marius and Velicu, Horia and Popescu, Marius and Dascalu, Mihai and Rebedea, Traian",
    booktitle = "Findings of the Association for Computational Linguistics: EMNLP 2024",
    month = nov,
    year = "2024",
    pages = "11632--11647"
}