--- license: apache-2.0 language: - ca - es - en tags: - valencian - spanish - english - text-generation - instruct - alia - gplsi datasets: - OpenAssistant/oasst2 - OpenAssistant/oasst1 - BSC-LT/m-personas - projecte-aina/RAG_Multilingual - facebook/flores - CohereLabs/aya_dataset - Unbabel/TowerBlocks-v0.2 - projecte-aina/MentorES - databricks/databricks-dolly-15k - tatsu-lab/alpaca - openai/gsm8k - Open-Orca/OpenOrca - HuggingFaceH4/no_robots - projecte-aina/CoQCat - gplsi/boua_parallel - allenai/scifact - somosnlp/LingComp_QA - somosnlp/instruct-legal-refugiados-es library_name: transformers pipeline_tag: text-generation --- # Aitana-2B-SI-Instruct **Aitana-2B-SI-Instruct** is an instruction-tuned generative language model from the **Aitana family**, developed by the [GPLSI (Language and Information Systems Group)](https://gplsi.dlsi.ua.es/) at the University of Alicante. Built on [gplsi/Aitana-2B-S-base-1.0](https://huggingface.co/gplsi/Aitana-2B-S-base-1.0), this model has been fine-tuned to follow instructions effectively across Valencian, Spanish, and English, with particular emphasis on enhancing Valencian language capabilities. ## Table of Contents - [Model Description](#model-description) - [Evaluation](#evaluation) - [Training Data](#training-data) - [Intended Uses](#intended-uses) - [How to Use](#how-to-use) - [Additional Information](#additional-information) ## Model Description | Property | Value | |----------|-------| | **Base Model** | [gplsi/Aitana-2B-S-base-1.0](https://huggingface.co/gplsi/Aitana-2B-S-base-1.0) | | **Architecture** | Transformer decoder-only | | **Parameters** | ~2.25B | | **Languages** | Valencian, Spanish, English | | **License** | Apache 2.0 | Aitana-2B-SI-Instruct is an instruction-tuned variant of Aitana-2B-S-base-1.0, fine-tuned on multilingual instruction data to follow user prompts and generate helpful responses across Valencian, Spanish, and English. The model was NOT instruction-tuned on Catalan data, though it retains some Catalan capabilities from its base model. ## Training Data This model was instruction fine-tuned on the ALIA Instruction/v12 dataset, composed of the following sources: | Dataset ID | Name | Languages | Source | |------------|------|-----------|--------| | ins1 | OpenAssistant2 (OASST2) | CA, EN, ES, VA | [OpenAssistant/oasst2](https://huggingface.co/datasets/OpenAssistant/oasst2) | | ins2 | OpenAssistant1 (OASST1) | CA, VA | [OpenAssistant/oasst1](https://huggingface.co/datasets/OpenAssistant/oasst1) | | ins3 | M-Personas | CA, EN, ES, VA | [BSC-LT/m-personas](https://huggingface.co/datasets/BSC-LT/m-personas) | | ins4 | RAG Multilingual | CA, EN, ES, VA | [projecte-aina/RAG_Multilingual](https://huggingface.co/datasets/projecte-aina/RAG_Multilingual) | | ins5 | FLORES | CA, EN, ES | [facebook/flores](https://huggingface.co/datasets/facebook/flores) | | ins6 | Aya Dataset | EN, ES, VA | [CohereLabs/aya_dataset](https://huggingface.co/datasets/CohereLabs/aya_dataset) | | ins7 | TowerBlocks | EN, ES | [Unbabel/TowerBlocks-v0.2](https://huggingface.co/datasets/Unbabel/TowerBlocks-v0.2) | | ins8 | Mentor / Mentores | CA, ES, VA | [projecte-aina/MentorES](https://huggingface.co/datasets/projecte-aina/MentorES) / [projecte-aina/MentorCA](https://huggingface.co/datasets/projecte-aina/MentorCA) | | ins9 | Dolly / Dolly 3K | CA, EN, VA | [databricks/databricks-dolly-15k](https://huggingface.co/datasets/databricks/databricks-dolly-15k) | | ins10 | Alpaca | EN, VA | [tatsu-lab/alpaca](https://huggingface.co/datasets/tatsu-lab/alpaca) | | ins11 | GSM8K | EN, VA | [openai/gsm8k](https://huggingface.co/datasets/openai/gsm8k) | | ins12 | OpenOrca | EN | [Open-Orca/OpenOrca](https://huggingface.co/datasets/Open-Orca/OpenOrca) | | ins13 | No Robots | EN | [HuggingFaceH4/no_robots](https://huggingface.co/datasets/HuggingFaceH4/no_robots) | | ins14 | CoQCA / CoQCat | CA, VA | [projecte-aina/CoQCat](https://huggingface.co/datasets/projecte-aina/CoQCat) | | ins15 | BOUA | ES | [gplsi/boua_parallel](https://huggingface.co/datasets/gplsi/boua_parallel) | | ins16 | SciFact | VA | [allenai/scifact](https://huggingface.co/datasets/allenai/scifact) | | ins17 | LingComp QA | VA | [somosnlp/LingComp_QA](https://huggingface.co/datasets/somosnlp/LingComp_QA) | | ins18 | Instruct Legal Refugiados | VA | [somosnlp/instruct-legal-refugiados-es](https://huggingface.co/datasets/somosnlp/instruct-legal-refugiados-es) | | ins19 | Amic-Paralelo | ES | [gplsi/amic_parallel](https://huggingface.co/datasets/gplsi/amic_parallel) | Catalan data was removed from the instruction tuning to focus on Valencian, Spanish, and English, though some Catalan appears in multilingual datasets. ## Intended Uses This model can be used for: - **Instruction following** in Valencian, Spanish, and English - **Chat and conversational applications** requiring multilingual support - **Text generation** with task-specific prompting - **Domain-specific applications** in administrative, legal, or tourism contexts > **Note**: As an instruction-tuned model, it is designed to follow user prompts and generate helpful responses. It is not intended for use as a factual knowledge base. ## How to Use ### Transformers ```python import torch from transformers import pipeline, AutoTokenizer model_id = "gplsi/Aitana-2B-SI-Instruct" tokenizer = AutoTokenizer.from_pretrained(model_id) generator = pipeline( "text-generation", model=model_id, tokenizer=tokenizer, torch_dtype=torch.bfloat16, device_map="auto", ) # Valencian example text = "Explica què són les Corts Valencianes i quina funció tenen." result = generator(text, do_sample=True, top_k=10, max_new_tokens=100) print(result[0]['generated_text']) # Spanish example text = "Describe las principales funciones del gobierno autonómico valenciano." result = generator(text, do_sample=True, top_k=10, max_new_tokens=100) print(result[0]['generated_text']) # English example text = "Explain the role of tourism in the Valencian Community economy." result = generator(text, do_sample=True, top_k=10, max_new_tokens=100) print(result[0]['generated_text']) ``` ## Evaluation In the following tables, we present the results obtained with different benchmarks from [lm-evaluation-harness](https://github.com/EleutherAI/lm-evaluation-harness) in comparison with [Salamandra-2B-Instruct](https://huggingface.co/BSC-LT/Salamandra-2B-Instruct). The results reflect the instruction-tuned performance of both models. ### Valencian #### Classification Benchmarks | Dataset | Lang. | Task | Metric | Salamandra-2B-Instruct | Aitana-2B-S-Instruct | |------------------------------|--------|----------------------------|-------------|---------------|-----------------------| | XNLI | va |Natural Language Inference | acc | **0.520** | 0.508 | #### Generation Benchmarks | Dataset | Lang. | Task | Metric | Salamandra-2B-Instruct | Aitana-2B-S-Instruct | |------------------------------|--------|----------------------------|-------------|---------------|-----------------------| | Cocoteros | va |Reading Comprehension | bleu | 2.796 | **3.121** | | Phrases ca-va | va-ca |Translation - Adaptation | bleu | 58.425 | **76.427** | | Phrases va-ca | va-ca |Translation - Adaptation | bleu | **70.660** | 68.902 | | Phrases va-es | va-es |Translation | bleu | 65.427 | **69.694** | | Phrases es-va | es-va |Translation | bleu | 45.688 | **55.725** | | Truthfulqa_va | va | Truthfulness | bleu_acc | 0.409 | **0.415** | ### Catalan #### Classification Benchmarks | Dataset | Lang. | Task | Metric | Salamandra-2B-Instruct | Aitana-2B-S-Instruct | |------------------------------|--------|---------------------------|-------------|---------------|-----------------------| | Belebele Cat_latn | ca | Reading Comprehension | acc | **0.287** | 0.274 | | COPA | ca | Commonsense Reasoning | acc | **0.708** | 0.704 | | XStoryCloze | ca | Commonsense Reasoning | acc | **0.616** | 0.615 | | OpenBookQA | ca | Question Answering | acc | 0.296 | **0.300** | | PAWS | ca | Paraphrasing | acc | 0.602 | **0.608** | | PiQA | ca | Question Answering | acc | 0.638 | **0.655** | | SiQA | ca | Question Answering | acc | **0.422** | 0.419 | | ARC Easy | ca | Question Answering | acc | 0.516 | **0.527** | | ARC Challenge | ca | Question Answering | acc | 0.298 | **0.305** | | XNLI | ca | Natural Language Inference| acc | 0.513 | **0.516** | | Teca | ca | Natural Language Inference| acc | 0.486 | **0.499** | | WNLI | ca | Natural Language Inference| acc | **0.563** | 0.451 | | Catcola | ca | Linguistic Acceptability | acc | 0.492 | **0.585** | | Catcola | ca | Linguistic Acceptability | mcc | **0.097** | -0.042 | | Catalanqa | ca | Question Answering | F1 | **0.516** | 0.397 | | Mgsm direct | ca | Math | exact match | 0.000 | **0.004** | | Catalanqa | ca | Question Answering | exact match | **0.182** | 0.029 | | Xquad | ca | Question Answering | exact match | **0.103** | 0.030 | | Xquad | ca | Question Answering | F1 | **0.394** | 0.303 | #### Generation Benchmarks | Dataset | Lang. | Task | Metric | Salamandra-2B-Instruct | Aitana-2B-S-Instruct | |------------------------------|--------|--------------------------|--------|----------------|-----------------------| | Cabreu abstractive | ca | Summarization | bleu | 7.610 | **9.199** | | Cabreu extractive | ca | Summarization | bleu | **38.002** | 14.869 | | Cabreu extreme | ca | Summarization | bleu | 2.733 | **4.209** | ### Spanish #### Classification Benchmarks | Dataset | Lang. | Task | Metric | Salamandra-2B-Instruct | Aitana-2B-S-Instruct | |------------------------------|--------|---------------------------|-------------|---------------|-----------------------| | Belebele | es | Reading Comprehension | acc | **0.268** | 0.260 | | PAWS | es | Paraphrasing | acc | 0.566 | **0.622** | | XNLI | es | Natural Language Inference| acc | **0.463** | 0.419 | | WNLI | es | Natural Language Inference| acc | 0.479 | **0.549** | | XStoryCloze | es | Commonsense Reasoning | acc | 0.617 | **0.619** | | Escola | es | Linguistic Acceptability | acc | 0.293 | **0.655** | | Escola | es | Linguistic Acceptability | mcc | 0.020 | **0.087** | | OpenbookQA | es | Question Answering | acc | 0.286 | **0.320** | | MGSM Direct | es | Math | exact match | 0.020 | **0.032** | | XQUAD | es | Question Answering | exact match | **0.066** | 0.039 | | XQUAD | es | Question Answering | F1 | **0.355** | 0.305 | #### Generation Benchmarks | Dataset | Lang. | Task | Metric | Salamandra-2B-Instruct | Aitana-2B-S-Instruct | |------------------------------|--------|---------------------|---------|----------------|-----------------------| | Cocoteros | es |Reading Comprehension| bleu | **3.308** | 2.508 | | XLSum | es | Summarization | bleu | 1.695 | **1.694** | ### English #### Classification Benchmarks | Dataset | Lang. | Task | Metric | Salamandra-2B-Instruct | Aitana-2B-S-Instruct | |------------------------------|--------|----------------------------|-------------|---------------|-----------------------| | Arc Challenge | en | Question Answering | acc | 0.354 | **0.362** | | Arc Easy | en | Question Answering | acc | 0.681 | **0.704** | | Belebele | en | Reading Comprehension | acc | 0.260 | **0.273** | | PAWS | en | Paraphrasing | acc | 0.597 | **0.610** | | XNLI | en | Natural Language Inference | acc | 0.512 | **0.553** | | XStoryCloze | en | Commonsense Reasoning | acc | **0.662** | 0.661 | | OpenBookQA | en | Question Answering | acc | 0.298 | **0.314** | | PiQA | en | Question Answering | acc | 0.715 | **0.720** | | Social iqa | en | Question Answering | acc | **0.453** | 0.431 | | WNLI | en | Natural Language Inference | acc | **0.535** | 0.451 | | MGSM Direct | en | Math | exact match | 0.008 | **0.088** | | TriviaQA | en | Question Answering | exact match | 0.076 | **0.170** | ### Judge Evaluation The model was also evaluated using an LLM-as-judge approach across different task categories. The scores below represent the average rating (1-5 scale, 5 being best) and standard deviation for each task category, comparing Aitana-2B-S-Instruct against Salamandra-2B-Instruct. | Task Category | Salamandra-2B-Instruct | Aitana-2B-S-Instruct | |---------------|------------------------|---------------------------| | CommonSense reasoning | **2.277 / 1.151** | 2.237 / 1.054 | | Maths | 1.060 / 0.124 | **1.105 / 0.209** | | Paraphrasing | **3.518 / 1.308** | 3.517 / 1.151 | | Reading comprehension | **2.966 / 1.111** | 2.740 / 1.348 | | Summarization | 2.217 / 1.068 | **2.267 / 0.967** | | Translation | **3.557 / 0.760** | 3.497 / 0.988 | | **Overall Avg** | **2.599 / 0.920** | 2.560 / 0.953 | ## Additional Information ### Author The model has been developed by the **Language and [Information Systems Group (GPLSI)](https://gplsi.dlsi.ua.es/)** and the **[Centro de Inteligencia Digital (CENID)](https://cenid.es)**, both part of the **[University of Alicante (UA)](https://www.ua.es/es/)**, as part of their ongoing research in **Natural Language Processing (NLP)**. ### 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*. This work has also been partially supported by Project HEART-NLP (PID2024-156263OB-C22). ### Acknowledgments We would like to express our gratitude to all individuals and institutions that have contributed to the development of this work. Special thanks to: - [Language Technologies Laboratory at Barcelona Supercomputing Center](https://www.bsc.es/es/discover-bsc/organisation/research-structure/language-technologies-laboratory) - [Centro Vasco de Tecnología de la Lengua (HiTZ)](https://www.hitz.eus/es) - [Centro Singular de Investigación en Tecnologías Inteligentes (CiTIUS)](https://citius.gal/) - [Sistemas Inteligentes de Acceso a la Información (SINAI)](https://www.ujaen.es/investigacion-y-transferencia/grupos-de-investigacion/sistemas-inteligentes-de-acceso-la-informacion-sinai) - [Instituto Universitario de Investigación Informática (IUII)](https://web.ua.es/es/iuii/) - [Leonardo HPC System](https://leonardo-supercomputer.cineca.eu/) - [European supercomputing ecosystem (EUROHPC)](https://www.eurohpc-ju.europa.eu/) 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**, whose contribution has been essential to the completion of this research. ### License [Apache License, Version 2.0](https://www.apache.org/licenses/LICENSE-2.0) ### Disclaimer This model is intended for general purposes and is available under a permissive Apache License 2.0. Be aware that the model may have biases and/or undesirable outputs. Users deploying systems based on this model are responsible for mitigating risks and complying with applicable AI regulations. ### Reference ```bibtex @misc{gplsi-Aitana-2B-SI-Instruct, author = {Martínez-Murillo, Iván and Sepúlveda-Torres, Robiert and Grande, Eduardo and Galiano, Santiago and Estevanell-Valladares, Ernesto L. and Consuegra-Ayala, Juan Pablo and Miró Maestre, María and Canal-Esteve, Miquel and Bonora, Mar and Gutierrez, Yoan and Abreu Salas, José Ignacio and Lloret, Elena and Montoyo, Andrés and Muñoz-Guillena, Rafael and Palomar, Manuel}, title = {Aitana-2B-SI-Instruct: Instruction-tuned model for Valencian, Spanish and English}, year = {2026}, institution = {Language and Information Systems Group (GPLSI) and Centro de Inteligencia Digital (CENID), University of Alicante (UA)}, howpublished = {\url{https://huggingface.co/gplsi/Aitana-2B-SI-Instruct}}, note = {Accessed: 2026-05-11} } ``` --- **Copyright © 2026 Language and Information Systems Group (GPLSI) and Centro de Inteligencia Digital (CENID), University of Alicante (UA). Distributed under the Apache License 2.0.**