README.md

---
license: apache-2.0
pipeline_tag: image-text-to-text
library_name: transformers
language:
- eu
- gl
- ca
- es
- en
datasets:
- HiTZ/latxa-corpus-v1.1
base_model:
- Qwen/Qwen3-VL-2B-Instruct
---

# Model Card for HiTZ/Latxa-Qwen3-VL-2B-Instruct

<p align="center">
  <img src="https://raw.githubusercontent.com/hitz-zentroa/latxa/refs/heads/main/assets/latxa_vision_circle.png" style="height: 350px;">
</p>

Latxa-Qwen3-VL-2B-Instruct is a Basque-adapted multimodal and multilingual instruct model built on top of Qwen3-VL-2B-Instruct, a powerful vision-language LLM capable of understanding and generating text and processing images. This model has been adapted by the HiTZ Research Center for improved performance on Basque (`mono_eu` variant), Galician and Catalan (`multi` variant) languages and interactive instruction following.

> [!WARNING]
> DISCLAIMER
> 
> These models are still under development.
> The released models are preliminary, and might be updated and improve in the future.

The released model contains several versions (revisions):
 - Multilingual (`multi`): in addition to Basque, the model has been also adapted to Galician and Catalan.
 - Basque monolingual (`mono_eu`): the Basque monolingual variant.

You can choose the model version by specifying the revision when loading the model with `revision="multi"`. By default (`main`) the multilingual variant is downloaded.

## Model Details

### Model Description

Latxa Vision models are a family of Vision-Language Models based on Qwen3-VL. The models were adapted to different languages following [Sainz et al. (2025)](https://aclanthology.org/2025.emnlp-main.1484/) adaptation method. The models are released under different language variants: `multi` (it has been adapted to Basque, Galician and Catalan) and `mono_eu` (adapted only to Basque). 

- **Developed by:** HiTZ Research Center & IXA Research group (University of the Basque Country UPV/EHU)
- **Funded by:** Ikergaitu and ALIA projects (Basque and Spanish Government)
- **Model type:** Vision-Language Instruct Model
- **Language(s) (NLP):** Basque, Galician, Catalan, Spanish, English and more.
- **License:** Apache 2.0
- **Finetuned from model:** Qwen3-VL-2B-Instruct

## Getting Started

Use the code below to get started with the model.

```python
from transformers import pipeline

# Load the text and image to text pipeline
pipe = pipeline("image-text-to-text", model="HiTZ/Latxa-Qwen3-VL-2B-Instruct", revision='multi')

# Messages can be of many types
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "image": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/cats.png"},
            {"type": "text", "text": "What do we see in this image?"},
        ]
    }
]
output = pipe(messages)
print(output)
```

## Uses

Latxa models are intended to be used with Basque data; for any other language the performance is not guaranteed.
Regarding the `multi` variant, it was additionally adapted for Galician and Catalan.


### Direct Use

Latxa Instruct models are trained to follow instructions or to work as chat assistants. 

### Out-of-Scope Use

The model is not intended for malicious activities, such as harming others or violating human rights. 
Any downstream application must comply with current laws and regulations. 
Irresponsible usage in production environments without proper risk assessment and mitigation is also discouraged. 


## Bias, Risks, and Limitations

In an effort to alleviate the potentially disturbing or harmful content, Latxa has been trained on carefully selected and processed 
data which comes mainly from local media, national/regional newspapers, encyclopedias and blogs (see [Latxa Corpus v1.1](https://huggingface.co/datasets/HiTZ/latxa-corpus-v1.1)). Still, the 
model is based on Qwen3-VL models and can potentially carry the same bias, risk and limitations.

### Recommendations

<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->

Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.

## Training Details

### Training Data

For training details, please, refer to our paper: [Instructing Large Language Models for Low-Resource Languages: A Systematic Study for Basque](https://aclanthology.org/2025.emnlp-main.1484/)


## Evaluation

We evaluated the models using 5-shot settings on multiple-choice and generative tasks.

| Task | Qwen3-VL 2B | 2B `mono_eu` | 2B `multi` | Qwen3-VL 4B | 4B `mono_eu` | 4B `multi` |
|------|:-----------:|:----------:|:-----------------:|:------------:|:----------:|:-----------------:|
| arc_eu_challenge_mc | 36.95 | 51.28 <span style="color:green">(+14.33)</span> | 55.20 <span style="color:green">(+18.25)</span> | 53.75 | 75.09 <span style="color:green">(+21.34)</span> | 75.34 <span style="color:green">(+21.59)</span> |
| arc_eu_easy_mc | 43.27 | 65.99 <span style="color:green">(+22.72)</span> | 69.95 <span style="color:green">(+26.68)</span> | 66.20 | 87.58 <span style="color:green">(+21.38)</span> | 87.58 <span style="color:green">(+21.38)</span> |
| belebele_eus_Latn | 46.00 | 65.44 <span style="color:green">(+19.44)</span> | 60.67 <span style="color:green">(+14.67)</span> | 69.67 | 80.67 <span style="color:green">(+11.00)</span> | 79.00 <span style="color:green">(+9.33)</span> |
| bertaqa_eu_global | 46.03 | 53.43 <span style="color:green">(+7.40)</span> | 56.81 <span style="color:green">(+10.78)</span> | 60.66 | 69.06 <span style="color:green">(+8.40)</span> | 69.65 <span style="color:green">(+8.99)</span> |
| bertaqa_eu_local | 37.27 | 42.51 <span style="color:green">(+5.24)</span> | 44.46 <span style="color:green">(+7.19)</span> | 40.27 | 53.43 <span style="color:green">(+13.16)</span> | 54.36 <span style="color:green">(+14.09)</span> |
| bl2mp | 49.11 | 87.94 <span style="color:green">(+38.83)</span> | 89.22 <span style="color:green">(+40.11)</span> | 55.89 | 90.17 <span style="color:green">(+34.28)</span> | 90.28 <span style="color:green">(+34.39)</span> |
| eus_exams_eu | 33.81 | 42.44 <span style="color:green">(+8.63)</span> | 42.81 <span style="color:green">(+9.00)</span> | 47.21 | 55.39 <span style="color:green">(+8.18)</span> | 56.40 <span style="color:green">(+9.19)</span> |
| eus_proficiency | 25.69 | 36.45 <span style="color:green">(+10.76)</span> | 36.58 <span style="color:green">(+10.89)</span> | 28.98 | 51.00 <span style="color:green">(+22.02)</span> | 51.77 <span style="color:green">(+22.79)</span> |
| eus_trivia | 35.04 | 40.41 <span style="color:green">(+5.37)</span> | 42.04 <span style="color:green">(+7.00)</span> | 44.49 | 56.27 <span style="color:green">(+11.78)</span> | 57.55 <span style="color:green">(+13.06)</span> |
| mgsm_native_cot_eu | 13.10 | 33.20 <span style="color:green">(+20.10)</span> | 34.00 <span style="color:green">(+20.90)</span> | 39.20 | 58.40 <span style="color:green">(+19.20)</span> | 62.40 <span style="color:green">(+23.20)</span> |
| mmlu_eu | 34.07 | 43.33 <span style="color:green">(+9.26)</span> | 45.93 <span style="color:green">(+11.86)</span> | 51.48 | 55.19 <span style="color:green">(+3.71)</span> | 57.41 <span style="color:green">(+5.93)</span> |
| piqa_eu_mc | 53.70 | 55.17 <span style="color:green">(+1.47)</span> | 54.08 <span style="color:green">(+0.38)</span> | 56.81 | 64.49 <span style="color:green">(+7.68)</span> | 68.68 <span style="color:green">(+11.87)</span> |
| siqa_eu_mc | 38.18 | 48.26 <span style="color:green">(+10.08)</span> | 50.31 <span style="color:green">(+12.13)</span> | 47.54 | 61.67 <span style="color:green">(+14.13)</span> | 62.59 <span style="color:green">(+15.05)</span> |
| xstorycloze_eu | 50.50 | 56.98 <span style="color:green">(+6.48)</span> | 57.05 <span style="color:green">(+6.55)</span> | 50.63 | 61.22 <span style="color:green">(+10.59)</span> | 61.81 <span style="color:green">(+11.18)</span> |
| **AVG EU** | **38.77** | **51.63 <span style="color:green">(+12.86)</span>** | **52.79 <span style="color:green">(+14.02)</span>** | **50.91** | **65.69 <span style="color:green">(+14.78)</span>** | **66.77 <span style="color:green">(+15.86)</span>** |



> [!WARNING]
> DISCLAIMER
> 
> These model are still under development.
> The results are only reported for Basque tasks, the results in the rest of the languages will be released in the near future.

## Citation

```bibtex
@inproceedings{sainz-etal-2025-instructing,
    title = "Instructing Large Language Models for Low-Resource Languages: A Systematic Study for {B}asque",
    author = "Sainz, Oscar  and
      Perez, Naiara  and
      Etxaniz, Julen  and
      Fernandez de Landa, Joseba  and
      Aldabe, Itziar  and
      Garc{\'i}a-Ferrero, Iker  and
      Zabala, Aimar  and
      Azurmendi, Ekhi  and
      Rigau, German  and
      Agirre, Eneko  and
      Artetxe, Mikel  and
      Soroa, Aitor",
    editor = "Christodoulopoulos, Christos  and
      Chakraborty, Tanmoy  and
      Rose, Carolyn  and
      Peng, Violet",
    booktitle = "Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing",
    month = nov,
    year = "2025",
    address = "Suzhou, China",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2025.emnlp-main.1484/",
    doi = "10.18653/v1/2025.emnlp-main.1484",
    pages = "29124--29148",
    ISBN = "979-8-89176-332-6",
    abstract = "Instructing language models with user intent requires large instruction datasets, which are only available for a limited set of languages. In this paper, we explore alternatives to conventional instruction adaptation pipelines in low-resource scenarios. We assume a realistic scenario for low-resource languages, where only the following are available: corpora in the target language, existing open-weight multilingual base and instructed backbone LLMs, and synthetically generated instructions sampled from the instructed backbone. We present a comprehensive set of experiments for Basque that systematically study different combinations of these components evaluated on benchmarks and human preferences from 1,680 participants. Our conclusions show that target language corpora are essential, with synthetic instructions yielding robust models, and, most importantly, that using as backbone an instruction-tuned model outperforms using a base non-instructed model. Scaling up to Llama 3.1 Instruct 70B as backbone, our model comes near frontier models of much larger sizes for Basque, without using any Basque instructions. We release code, models, instruction datasets, and human preferences to support full reproducibility in future research on low-resource language adaptation."
}
```

## Acknowledgements

This work has been partially supported by the
Basque Government (Research group funding
IT1570-22 and IKER-GAITU project), the Span-
ish Ministry for Digital Transformation and of
Civil Service, and the EU-funded NextGenera-
tionEU Recovery, Transformation and Resilience
Plan (ALIA project). The models were trained on the
Leonardo supercomputer at CINECA under the
EuroHPC Joint Undertaking, project EHPC-EXT-2024E01-042.