gplsi/oasst2_va

🗣️ OPENASSISTANTVA Dataset

This dataset is a Valencian translation subset of the OpenAssistant/oasst2 dataset, created by carefully selecting Spanish-language messages and translating them into Valencian.

This dataset supports research and development of large language models (LLMs) and conversational AI systems in Valencian, contributing to the availability of high-quality aligned conversational data for under-resourced language varieties.

📖 Origin

• —Source dataset: OpenAssistant/oasst2
• —Selection criteria: Spanish-language messages
• —Processing: Manual translation into Valencian by human language experts
• —Quality control: Expert-reviewed translations to ensure linguistic accuracy and naturalness

The goal was to build a high-quality alignment-oriented dataset in Valencian, preserving behavioral and safety-related characteristics from the original corpus.

🔎 Selection Process

The subset was created through a two-stage process.

First, conversations were filtered from the original Oasst2 dataset based on language, quality ranking, and non-synthetic origin. The following automatic filtering criteria were applied:

• —Language: Spanish
• —Rank: 0–2
• —Synthetic: false (only non-synthetic data)

Second, the selected conversations were manually analyzed and grouped into four behavioral categories reflecting assistant alignment patterns:

• —Affirmations
• —Uncertainty and controversial topics
• —Existential and capability-related questions
• —Refusal and ethical constraints

This categorization ensured the dataset captures a diverse range of assistant behaviors relevant to LLM alignment and safety research.

📂 Dataset Structure

Each row corresponds to a single message within a conversation tree and preserves much of the original OpenAssistant metadata.

Core Fields

Other Fields

Moderation & Review Metadata, Safety & Toxicity Signals, and Annotation & Feedback Signals.

🎯 Alignment Use Cases

OPENASSISTANTVA is particularly suited for:

• —Supervised Fine-Tuning (SFT) in Valencian
• —Refusal behavior training
• —Safety-aware LLM alignment
• —Cross-lingual alignment transfer
• —Multilingual RLHF experimentation
• —Evaluation of aligned behavior in low-resource languages

The dataset preserves refusal patterns, uncertainty calibration, and ethical boundary behaviors critical for modern LLM alignment pipelines.

🌍 Language

• —Valencian
• —Human expert translation from Spanish
• —Pragmatic intent and alignment behavior preserved

⚠️ Notes

• —This is a translated alignment-focused subset, not the full oasst2 dataset.
• —Only originally Spanish, non-synthetic messages were included.
• —Manual translation ensures high linguistic quality.
• —Metadata from the original dataset has been retained for research reproducibility.

💰 Funding

This work is funded by the Ministerio para la Transformación Digital y de la Función Pública, co-financed by the EU – NextGenerationEU, within the framework of the project Desarrollo de Modelos ALIA.

🙏 Acknowledgments

We acknowledge the creators and contributors of the original OpenAssistant/oasst2 dataset.

Also, we would like to express our gratitude to all individuals and institutions that have contributed to the development of this work.

<!--

Special thanks to:

• —[Data providers]
• —[Technological support providers]

We also acknowledge the financial, technical, and scientific support of the Ministerio para la Transformación Digital y de la Función Pública - Funded by EU – NextGenerationEU within the framework of the project Desarrollo de Modelos ALIA. -->

📚 Reference

Please cite the dataset using the following BibTeX entry:

@misc{oasst2_va_2026,
  author       = {Picazo-Izquierdo, Alicia and Consuegra-Ayala, Juan Pablo and Muñoz Guillena, Rafael},
  title        = {OPEN_ASSISTANT_VA Dataset,
  year         = {2026},
  institution  = {Language and Information Systems Group (GPLSI) and Centro de Inteligencia Digital (CENID), University of Alicante (UA)},
  howpublished = {\url{https://huggingface.co/datasets/gplsi/oasst2_va}},
}

⚠️ Disclaimer

When third parties deploy systems or provide services based on this data , or use the data themselves, they bear the responsibility for mitigating any associated risks and ensuring compliance with applicable regulations, including those governing the use of Artificial Intelligence.

The University of Alicante, as the owner and creator of the dataset, shall not be held liable for any outcomes resulting from third-party use.

📜 License

This dataset is released under the:

Apache License 2.0

https://www.apache.org/licenses/LICENSE-2.0