---
language:
  - ro
license: cc-by-nc-4.0
library_name: transformers
tags:
  - gemma3
  - gemma
  - romanian
  - vlm
  - instruct
  - multimodal
datasets:
  - 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
base_model:
  - google/gemma-3-4b-it
model-index:
- name: OpenLLM-Ro/RoGemma3-4B-Instruct
  results:
  - task:
      type: image-text-to-text
    dataset:
      name: Romanian_VLM_Benchmarks
      type: Romanian_VLM_Benchmarks
    metrics:
    - name: Micro avg.
      type: Score
      value: 59.54
    - name: Macro avg.
      type: Score
      value: 57.14
  - task:
      type: image-text-to-text
    dataset:
      name: MMBench
      type: MMBench
    metrics:
    - name: Accuracy
      type: accuracy
      value: 69.96
  - task:
      type: image-text-to-text
    dataset:
      name: MMStar
      type: MMStar
    metrics:
    - name: Accuracy
      type: accuracy
      value: 46.01
  - task:
      type: image-text-to-text
    dataset:
      name: SeedBench2
      type: SeedBench2
    metrics:
    - name: Accuracy
      type: accuracy
      value: 62.83
  - task:
      type: image-text-to-text
    dataset:
      name: MMMU
      type: MMMU
    metrics:
    - name: Accuracy
      type: accuracy
      value: 38.67
  - task:
      type: image-text-to-text
    dataset:
      name: MME
      type: MME
    metrics:
    - name: Accuracy
      type: accuracy
      value: 54.62
  - task:
      type: image-text-to-text
    dataset:
      name: CVQA
      type: CVQA
    metrics:
    - name: Accuracy
      type: accuracy
      value: 64.24
  - task:
      type: image-text-to-text
    dataset:
      name: ALM-Bench
      type: ALM-Bench
    metrics:
    - name: Score
      type: score
      value: 65.40
  - task:
      type: image-text-to-text
    dataset:
      name: RoMemes
      type: RoMemes
    metrics:
    - name: F1
      type: f1
      value: 40.78
  - task:
      type: image-text-to-text
    dataset:
      name: RoCultVLM
      type: RoCultVLM
    metrics:
    - name: Score
      type: score
      value: 52.84
  - task:
      type: image-text-to-text
    dataset:
      name: RoFlickr30k-Caption
      type: RoFlickr30k-Caption
    metrics:
    - name: BERTScore
      type: bertscore
      value: 84.35
  - task:
      type: image-text-to-text
    dataset:
      name: RoFlickr30k-QA
      type: RoFlickr30k-QA
    metrics:
    - name: Score
      type: score
      value: 84.74
  - task:
      type: image-text-to-text
    dataset:
      name: LLaVA-Wild
      type: LLaVA-Wild
    metrics:
    - name: Score
      type: score
      value: 55.71
  - task:
      type: image-text-to-text
    dataset:
      name: AyaVisionBench
      type: AyaVisionBench
    metrics:
    - name: Score
      type: score
      value: 47.04
  - task:
      type: image-text-to-text
    dataset:
      name: m-WildVision
      type: m-WildVision
    metrics:
    - name: Score
      type: score
      value: 57.60
  - task:
      type: image-text-to-text
    dataset:
      name: RoCosyn
      type: RoCosyn
    metrics:
    - name: Score
      type: score
      value: 59.06
  - task:
      type: image-text-to-text
    dataset:
      name: RoFinepdfs
      type: RoFinepdfs
    metrics:
    - name: ANLS
      type: anls
      value: 84.35
  - task:
      type: image-text-to-text
    dataset:
      name: RoMemes OCR
      type: RoMemes-OCR
    metrics:
    - name: ANLS
      type: anls
      value: 86.33
  - task:
      type: image-text-to-text
    dataset:
      name: PixmoCount
      type: PixmoCount
    metrics:
    - name: Exact match
      type: exact_match
      value: 51.80
  - task:
      type: image-text-to-text
    dataset:
      name: PixmoPoints
      type: PixmoPoints
    metrics:
    - name: F1
      type: f1
      value: 24.87
---

# 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](https://huggingface.co/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](https://huggingface.co/google/gemma-3-4b-it)
- **Trained using:**
  - [OpenLLM-Ro/ro_sft_laion](https://huggingface.co/datasets/OpenLLM-Ro/ro_sft_laion)
  - [OpenLLM-Ro/ro_sft_pixmo_cap](https://huggingface.co/datasets/OpenLLM-Ro/ro_sft_pixmo_cap)
  - [OpenLLM-Ro/ro_sft_flickr30k_cap](https://huggingface.co/datasets/OpenLLM-Ro/ro_sft_flickr30k_cap)
  - [OpenLLM-Ro/ro_sft_llava_mix](https://huggingface.co/datasets/OpenLLM-Ro/ro_sft_llava_mix)
  - [OpenLLM-Ro/ro_sft_pixmo_aa](https://huggingface.co/datasets/OpenLLM-Ro/ro_sft_pixmo_aa)
  - [OpenLLM-Ro/ro_sft_pixmo_cap_qa](https://huggingface.co/datasets/OpenLLM-Ro/ro_sft_pixmo_cap_qa)
  - [OpenLLM-Ro/ro_sft_flickr30k_qa](https://huggingface.co/datasets/OpenLLM-Ro/ro_sft_flickr30k_qa)
  - [OpenLLM-Ro/ro_sft_cosyn](https://huggingface.co/datasets/OpenLLM-Ro/ro_sft_cosyn)
  - [OpenLLM-Ro/ro_sft_finepdfs](https://huggingface.co/datasets/OpenLLM-Ro/ro_sft_finepdfs)
  - [OpenLLM-Ro/ro_sft_pixmo_points](https://huggingface.co/datasets/OpenLLM-Ro/ro_sft_pixmo_points)
  - [OpenLLM-Ro/ro_sft_pixmo_count](https://huggingface.co/datasets/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

```python
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

| Model | Micro avg. | Macro avg. |
|---|---|---|
| Gemma3-4B-it | 55.53 | 52.36 |
| RoGemma3-4B-Instruct | **59.54** | **57.14** |

### General Understanding

| Model | MMBench | MMStar | SeedBench2 |
|---|---|---|---|
| Gemma3-4B-it | 59.13 | 41.49 | 57.55 |
| RoGemma3-4B-Instruct | **69.96** | **46.01** | **62.83** |

### Knowledge & Reasoning

| Model | MMMU | MME |
|---|---|---|
| Gemma3-4B-it | 36.67 | **57.32** |
| RoGemma3-4B-Instruct | **38.67** | 54.62 |

### Cultural

| Model | CVQA | ALM-Bench | RoMemes | RoCultVLM |
|---|---|---|---|---|
| Gemma3-4B-it | **64.90** | **65.97** | **43.24** | 52.19 |
| RoGemma3-4B-Instruct | 64.24 | 65.40 | 40.78 | **52.84** |

### Generation & Open-ended

| Model | RoFlickr30k-Caption | RoFlickr30k-QA | LLaVA-Wild | AyaVisionBench | m-WildVision |
|---|---|---|---|---|---|
| Gemma3-4B-it | 70.93 | 81.66 | 54.84 | **52.81** | **60.80** |
| RoGemma3-4B-Instruct | **84.35** | **84.74** | **55.71** | 47.04 | 57.60 |

### OCR & Documents

| Model | RoCosyn | RoFinepdfs | RoMemes OCR |
|---|---|---|---|
| Gemma3-4B-it | 48.40 | 67.12 | **89.47** |
| RoGemma3-4B-Instruct | **59.06** | **84.35** | 86.33 |

### Grounding

| Model | PixmoCount | PixmoPoints |
|---|---|---|
| Gemma3-4B-it | 40.42 | 10.21 |
| RoGemma3-4B-Instruct | **51.80** | **24.87** |

## Citation

```bibtex
@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"
}
```