BSC-LT / ALIA-40b-instruct

WARNING: ALIA-40b-Instruct is an instruction-tuned model with a preliminary alignment process. It has not yet undergone a full alignment procedure to ensure safety. The model may generate biased, factually incorrect, harmful, or inappropriate content. Users should refer to the Limitations section and apply additional filtering and alignment processes before deploying this model in production.

ALIA-40b-instruct Model Card

The ALIA-40b-instruct model is an instructed variant of a context-extended base ALIA-40b model , which was pre-trained from scratch on 9.83 trillion tokens of carefully curated data spanning 35 European languages (including code). This instructed version is optimized to follow user prompts and engage in dialogue. It supports a broad range of languages (e.g. Spanish, Catalan, Basque, English, etc.) and is capable of text generation, translation, summarization, and question-answering in these languages. This version has also gone through a preliminary alignment phase for helpfulness and safety with synthetically generated preference pairs.

In keeping with our commitment to open-source development, all tools and sources used to process and create the training data are open-licensed. For clarity, our definition of open-licensed excludes any source, tool, model, or dataset whose terms of use impose restrictive conditions that impede standard open reuse.

This model is released under the permissive Apache 2.0 license . Along with the open weights, all training scripts and configuration files are made publicly available in this GitHub repository .

To visit the model cards of other model versions, please refer to the Model Index .

---

Model Details

Description

The ALIA-40b is a transformer-based, decoder-only language model that was pre-trained from scratch on 9.37 trillion tokens of meticulously curated data. It subsequently underwent continued pretraining on additional 424 billion high-quality tokens, and was further extended with a supplementary 39 billion tokens drawn from a similarly diverse mixture, totalling 9.83 trillion tokens.

ALIA-40b-Instruct is an instructed variant of this latest ALIA-40b version. Its post-training process comprises three consecutive stages, each targeting a specific capability: (1) long-context adaptation to extend the model’s context window, (2) supervised fine-tuning to improve instruction following capabilities, and (3) a preliminary alignment stage to better match human preferences and safety.

After the long-context adaptation, the model enters the supervised fine-tuning (SFT) stage. This stage is implemented in two phases for efficiency reasons: a short-context SFT with 469k conversation examples to strengthen instruction following, followed by a long-context SFT with 9k long-context instances. We separate these phases because full-context fine-tuning is computationally expensive.

In the third stage, the model is aligned with human preferences through Direct Policy Optimization (DPO) using a mixture of 403k preference pairs. Of this mixture, approximately 82% of the pairs target general model helpfulness, while 18% focus on response safety. This alignment stage is preliminary, and further work is ongoing to strengthen safety and reliability.

Although the base model is highly multilingual, the post-training process concentrated primarily on Spanish, Catalan, Basque, Galician, and English. We also incorporated data from other related languages where inclusion empirically improved the performance on the target languages. However, performance in those additional languages is not guaranteed due to the limited amount of available data and the scarcity of evaluation resources.

Hyperparameters

Here we list the specific hyperparameters used during the different training stages.

Long context CPT

Short context SFT

Long context SFT

Alignment

Architecture

---

Intended Use

Direct Use

ALIA‑40b‑instruct is intended for research and development purposes as a general-purpose multilingual assistant. It can be used to generate text, answer questions, translate between supported languages, and follow user instructions in those languages. As noted by the ALIA-40b base card, the ALIA family is aimed at both research and commercial use in any of the covered languages. In practice, ALIA-40b-instruct is best suited for tasks like multilingual chatbots, summarization, translation, and content generation, provided users are aware of its limitations.

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.

---

Hardware and Software

Training Framework

The post-training process was conducted using three complementary frameworks, each selected to best support its corresponding stage:

• Supervised Fine-Tuning (SFT): Conducted with an internal fork of the FastChat codebase, adapted to our infrastructure and optimized for stability and efficiency in our use case.
• Long-Context SFT: Performed using NeMo-Aligner, chosen to ensure compatibility with extended-context training while maintaining consistency with the FastChat-based SFT.
• Alignment Stage: Implemented with the TRL (Transformers Reinforcement Learning) library, applied to preference-pair training to achieve preliminary alignment with human preferences.

Compute Infrastructure

All models were trained on MareNostrum 5 , a pre-exascale EuroHPC supercomputer hosted and operated by Barcelona Supercomputing Center.

The accelerated partition is composed of 1,120 nodes with the following specifications:

• 4x Nvidia Hopper GPUs with 64GB HBM2 memory
• 2x Intel Sapphire Rapids 8460Y+ at 2.3Ghz and 32c each (64 cores)
• 4x NDR200 (BW per node 800Gb/s)
• 512 GB of Main memory (DDR5)
• 460GB of NVMe storage

The table below specifies the number of nodes and GPUs employed for each post-training stage:

---

How to use

The instruction-following models utilize the widely adopted ChatML template to structure conversational inputs and outputs.

Using this standardized chat format ensures a consistent and enhanced conversational experience. The template can be easily applied through the tokenizer’s built-in functions, as illustrated in the example snippet below:

from transformers import AutoTokenizer, AutoModelForCausalLM
import transformers
import torch

model_id = "BSC-LT/ALIA-40b-instruct"

text = "At what temperature does water boil?"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    device_map="auto",
    torch_dtype=torch.bfloat16
  )

message = [ { "role": "user", "content": text } ]

prompt = tokenizer.apply_chat_template(
    message,
    tokenize=False,
    add_generation_prompt=True,
)

inputs = tokenizer.encode(prompt, add_special_tokens=False, return_tensors="pt")
outputs = model.generate(input_ids=inputs.to(model.device), max_new_tokens=200)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

---

Using this template, each turn in the conversation is preceded by a <|im_start|> delimiter indicating the beginning of a message, followed by the role of the entity (either user , for content supplied by the user, or assistant for the model's responses), and finished with the <|im_end|> token:

<s><|im_start|>user
At what temperature does water boil?<|im_end|>
<|im_start|>assistant
Water turns into vapor at 100°C.<|im_end|>

---

Instruction Tuning Data

The dataset used in the initial supervised fine-tuning stage consists of 469k conversations, each with a maximum token length of 4k. The training mixture is obtained by combining a selection of (human and synthetic) permissive-licensed datasets, with a collection of synthetic conversations curated in-house .

The synthetic conversations are generated using DeepSeek-V3-0324 , leveraging seed data and prompts from pre-training corpora, as well as other openly available instruction datasets.

The table below provides a detailed breakdown of the datasets included in this mixture, specifying their origin, type, license, and contribution to the overall corpus:

Following the short-context supervised fine-tuning, a second stage was introduced using the remaining 9k short-context samples from our mix, together with 480 long-context samples.

The long-context data was synthetically generated with Salamandra-7B using source texts from FineWebEdu, FineWeb2, and Wikipedia. The length of the examples varies between 16k and 160k tokens. The resulting outputs were subsequently filtered with the same DeepSeek-V3-0324 model to ensure quality and consistency.

The table below summarizes the distribution of instructions by language included in the long-context supervised fine-tuning stage:

Detailed SFT Data Sources:

The following table provides a detailed overview of the supervised fine-tuning data sources, including the dataset name, generation method, license and a brief description of each:

SFT Datasets

Dataset	Generation Method	License	Description
aya-dataset	Human Crowdsourced	Apache-2.0	aya_dataset for the languages of interest.
coqcat-train	Human Annotation	CC-BY-NC-ND-4.0	CoQCat train split, formatted using conversational templates.
databricks-dolly-15k	Human Annotation	CC-BY-SA-3.0	databricks-dolly-15k dataset.
dolly-ca	Human Translation	CC-BY-SA-3.0	dolly3k_ca dataset.
flores-dev	Human	CC-BY-SA-4.0	Flores-200 dev split, formatted using conversational templates.
mentor-es	Human Annotation	CC-BY-4.0	MentorES dataset.
mentor-ca	Machine Translation	CC-BY-4.0	MentorCA dataset. Machine translated version of MentorES.
no-robots	Human Annotation	CC-BY-NC-4.0	no_robots dataset.
rag-multilingual	Synthetic	CC-BY-SA-4.0	RAG_Multilingual dataset. Synthetic QA dataset generated with Mixtral8x7b.
tower-blocks	Mixture	Various licenses (only open licensed instances are used)	TowerBlocks-v0.2 filtered by subdataset license and the languages of interest.
oasst2_self-identity-rephrase	Human Crowdsourced / Synthetic	Apache-2.0	oasst2 dataset for the languages of interest. Subsequently rephrased to adapt the model’s identity information to our case using DeepSeek-V3-0324.
self-identity	Synthetic	Apache-2.0 (internal)	Conversations involving self-identity information of the model, synthetically curated using DeepSeek-V3-0324.
open-r1-math	Synthetic	Apache-2.0	Default 93k split of the OpenR1-Math-220k dataset.
open-r1-math_translated	Synthetic	Apache-2.0 (internal)	OpenR1-Math-220k default split translated to the languages of interest with DeepSeek-V3-0324.
fineweb-edu_qa	Synthetic	Apache-2.0 (internal)	QA conversations created by prompting DeepSeek-V3-0324 with the highest quality documents of FineWeb-Edu . Subsequently filtered with the same model to ensure self-contained question-answering pairs meet quality thresholds.

*All externally sourced datasets have undergone a sanity check using shallow rule-based filtering to discard incorrect or low-quality samples and ensure conversational quality.*

Alignment Data

The alignment data was synthetically generated from a corpus of approximately 403k prompts designed to improve both helpfulness and safety.

• Helpfulness : Prompts include instruction following, mathematics, question answering, and reasoning tasks across Catalan, Spanish, English, Euskera, and Galician. Additionally, M-Personas conversations, a resource specifically generated for this project, were incorporated and will also be released.
• Safety : Prompts were synthetically generated from seed prompts written by human annotators, covering nine harm categories to ensure broad coverage of safety-related scenarios.

Following approaches similar to UltraFeedback and PKU, each instruction underwent the following process:

1. Multiple responses were produced using a pool of permissively licensed models (see Model Pool ) on helpfulness or safety, depending on the prompt.
2. These responses were rated by a judge (Deepseek-V3-0324). Helpfulness responses were given an overall rating, while safety responses were given a score based on their level of severity over a list of harm categories.
3. Preference pairs were constructed from these ratings. This phase should be considered preliminary, as future versions of the model will incorporate human annotators to refine and curate the generation and evaluation pipeline.

The table below presents the distribution of helpfulness prompts by language, detailing the number of examples contributed from each language:

The following table summarizes the safety prompts included in the alignment dataset by language and number of instances, covering the nine harm categories:

Model Pool for Synthetic Data Generation

In the table below, we list the permissively licensed models that were used to generate the synthetic datasets for alignment:

Model Pool

Family	Model Name	Size (B)	Variant	License
EuroLLM	EuroLLM_9B_Instruct	9	instructed	Apache 2.0
Deepseek	DeepSeek-V3-0324	685	aligned	MIT
Qwen	Qwen3-235B-A22B	235	aligned	Apache 2.0
	Qwen3-30B-A3B	30	aligned	Apache 2.0
	Qwen3-32B	32	aligned	Apache 2.0
	Qwen3-14B	14	aligned	Apache 2.0
	Qwen3-8B	8	aligned	Apache 2.0
Mistral	Mixtral-8x7B-Instruct-v0.1	56	aligned	Apache 2.0
	Mistral-7B-Instruct-v0.3	7	aligned	Apache 2.0
	Mistral-Small-24B-Instruct-2501	24	aligned	Apache 2.0
	Mistral-Nemo-Instruct-2407	12	instructed	Apache 2.0
OLMO	OLMo-2-0325-32B-SFT	32	instructed	Apache 2.0
	OLMo-2-1124-13B-SFT	13	instructed	Apache 2.0
	OLMo-2-1124-7B-SFT	7	instructed	Apache 2.0
FLOR_BSC	Aitana_6_3B_BSC_Instructed	6.3	instructed	Apache 2.0
	Flor_6_3B_Instruct	6.3	instructed	Apache 2.0
Salamandra	Salamandra-40b_pre-1.0_sft-1.0_hh_rlhf_ali	40	instructed	Apache 2.0
	Salamandra-40b_pre-1.0_sft-1.0_hh_rlhf_tox	40	instructed	Apache 2.0
	Salamandra-2b_pre-1.2_sft-1.0_hh_rlhf_ali	2	instructed	Apache 2.0
	Salamandra-7b_pre-1.2_sft-1.0_hh_rlhf_ali	7	instructed	Apache 2.0
	Salamandra-2b_pre-1.2_sft-1.0_hh_rlhf_tox	2	instructed	Apache 2.0
	Salamandra-7b_pre-1.2_sft-1.0_hh_rlhf_tox	7	instructed	Apache 2.0

Evaluation

Gold-standard benchmarks

Evaluation is done using the Language Model Evaluation Harness (Gao et al., 2024). We evaluate on a set of tasks taken from SpanishBench , CatalanBench , BasqueBench and GalicianBench , as well as existing English tasks available in the LM Evaluation Harness. These benchmarks include both new and existing tasks and datasets. The tables below report results for a representative selection of evaluation datasets, capturing model's performance across a variety of tasks within these benchmarks.

Only tasks that are human-generated, human-translated, or involve strong human-in-the-loop process (i.e., machine translation followed by professional revision or machine generation followed by human revision and annotation) were used. This approach explains the variation in the number of tasks reported across languages. As additional high-quality tasks are published, we will update the evaluation results accordingly. We also plan to expand evaluation to other languages, provided that the datasets meet our quality standards.

During the implementation of the evaluation we observed a series of issues worth considering when replicating and interpreting the results presented. These issues include ≈1.5% variances in performance in some tasks depending on the version of the transformers library used, and depending on the use (or lack of use) of tensor parallelism when loading a model. When implementing existing tasks, we carry out a comprehensive quality evaluation of the dataset, the Harness task itself, and what kind of input models see during evaluation. Our implementation (see links above) addresses multiple existing problems such as errors in datasets and prompts, and lack of pre-processing. All this means that results will vary if using other Harness implementations, and may slightly vary depending on the replication setup.

It should be noted that these results are subject to all the drawbacks of every current gold-standard evaluation, and that the figures do not fully represent the model's capabilities and potential. We thus advise caution when reading and interpreting the results.

All results reported below correspond to a 0-shot evaluation setting.

Spanish

Catalan

Basque

Galician

English

Current LM Evaluation Harness implementation is lacking correct pre-processing. These results are obtained with adequate pre-processing.

LLM-as-a-judge

We use Prometheus-2 8x7B as a judge to evaluate the responses of the model. Tasks are created from existing multilingual evaluation datasets covering the same categories as the ones measured in our gold-standard benchmarks. We randomly select a subset of 250 instances per language from the test set of each source dataset. To evaluate the responses of our model, we use task-specific criteria developed in-house for the LLM-judge to use. Each criterion is measured either as a 5-point Likert scale or as a binary task depending on the idiosyncrasy of the task and criterion.

Prompts for each task are created in various ways to score the model's robustness in addition to these criteria. This is done by presenting the same source instance within three different prompts. We then calculate the variance between the scores assigned by the LLM-judge to our model's responses to the three prompt styles and average it across all instances. Prompts are human translated to all languages measured. We do not provide the LLM-judge with a reference answer.

The judge prompt we use during evaluation is the same used to fine tune the Prometheus-2 family. We keep the judge prompt and criteria used to present the LLM-judge with the task prompts and model responses in English for evaluation across languages. The judge prompt used is:

"You are a fair judge assistant tasked with providing clear, objective feedback based on specific criteria, ensuring each assessment reflects the absolute standards set for performance.

###Task Description:
An instruction (might include an Input inside it), a response to evaluate, and a score rubric representing a evaluation criteria are given.
1. Write a detailed feedback that assess the quality of the response strictly based on the given score rubric, not evaluating in general.
2. After writing a feedback, write a score that is an integer between {a} and {b}. You should refer to the score rubric.
3. The output format should look as follows: \"Feedback: (write a feedback for criteria) [RESULT] (an integer number between {a} and {b})\"
4. Please do not generate any other opening, closing, and explanations.

###The instruction to evaluate:
{input}

###Response to evaluate:
{prediction}

###Score Rubrics:
{criteria}

###Feedback:"

As an example, prompts for the Math task in English are based on instances from MGSM , and each instance is presented within these prompts:

"en": [
      ("I need help with this math problem: \"", "\" Give me the answer step by step and also the final result separately."),
      ("Can you please help me answer this? \"", "\" Explain the answer and give me the final result as well. Thanks."),
      ("Help me with this problem: \"", "\" I need the answer explained and the final result separately.")
]

This task is then evaluated by the LLM-judge using two criteria, reasoning capability (5-point Likert) and mathematical correctness (binary):

reasoning_capability_criteria = {
    "reasoning_capability": """
[Does the model's answer demonstrate reasoning capability?]
Score 1: The answer demonstrates poor reasoning, with illogical arguments or conclusions that do not follow from the provided information.
Score 2: The answer shows weak reasoning, with some logical connections but also contains significant flaws or gaps in the argumentation.
Score 3: The answer demonstrates adequate reasoning, with generally logical arguments, but may have minor flaws or a lack of depth in the reasoning process.
Score 4: The answer shows strong reasoning, with well-structured arguments and conclusions that logically follow from the information provided.
Score 5: The answer demonstrates exceptional reasoning, with clear, coherent, and insightful arguments that are logically sound and well-supported by the information provided."""
}

mathematical_correctness_binary_criteria = {
    "mathematical_correctness_binary": """
[Is the model's answer mathematically correct?]
Score 0: The answer contains mathematical errors that render the solution incorrect or unreliable.
Score 1: The answer is mathematically correct, with accurate calculations and appropriate use of mathematical concepts."""
}

Multilingual results

Here, we present results for seven categories of tasks in Spanish, Catalan, Basque, Galician, and English. Results are presented for each task, criterion and language. Criteria with a (B) after their name are binary criteria (i.e., numbers go from 0 to 1, where 1 is best). The rest of the criteria are measured using a 5-point Likert scale, where 5 is best. The first number of the pair of numbers separated by / shows the average score for the criterion (and language). The second number of each pair is the robustness score, where numbers closer to 0 mean that the model generates similar responses when comparing the three prompt varieties for a single instance.

Category	Dataset	Criteria	es	ca	gl	eu	en
Commonsense Reasoning	XStoryCloze	Ending Coherence	3.17/0.65	3.19/0.46	2.87/0.57	1.94/0.49	3.62/0.51
Paraphrasing	PAWS	Completeness (B)	0.78/0.10	0.66/0.14	0.73/0.13	0.53/0.14	0.76/0.11
Paraphrase Generation	3.52/0.76	3.31/0.89	3.28/0.85	2.78/0.95	3.53/0.61
Grammatical Correctness (B)	0.88/0.06	0.83/0.09	0.84/0.08	0.75/0.12	0.90/0.06
Reading Comprehension	Belebele	Answer Relevance (B)	0.82/0.06	0.83/0.06	0.80/0.07	0.65/0.11	0.82/0.07
Passage Comprehension	3.25/0.44	3.25/0.45	3.15/0.59	2.52/0.46	3.27/0.45
Extreme Summarization	XLSum & caBreu & summarization_gl	Informativeness	3.49/0.23	3.60/0.18	3.52/0.19	--/--	3.37/0.26
Conciseness	3.28/0.20	3.28/0.20	3.39/0.19	--/--	3.37/0.22
Mathematics	mgsm	Mathematical Correctness (B)	0.81/0.10	0.83/0.09	0.82/0.09	0.90/0.06	0.76/0.11
Reasoning Capability	3.74/0.72	3.69/0.54	3.61/0.56	3.46/0.39	3.48/0.77
Translation form Language	FLoRes	Accuracy	4.00/0.16	4.13/0.16	4.09/0.15	3.73/0.20	4.21/0.17
Fluency	3.67/0.13	3.76/0.12	3.83/0.11	3.37/0.14	3.79/0.13
Translation to Language	FLoRes	Accuracy	4.06/0.19	4.06/0.14	3.88/0.16	3.61/0.19	4.47/0.16
Fluency	3.78/0.13	3.74/0.12	3.50/0.15	3.18/0.11	4.06/0.13

Long Context Evaluation

To assess the long-context capabilities of our model, we performed a "needle in a haystack" test with the following configuration:

• Needle Phrase : "The best thing to do in San Francisco is eat a sandwich and sit in Dolores Park on a sunny day."
• System Prompt: “You are a helpful AI bot that answers questions for a user. Keep your response short and direct”
• Retrieval Question : "What is the best thing to do in San Francisco?"
• Evaluator : prometheus-8x7b-v2.0 , used as the evaluation judge to determine whether the model correctly retrieved and utilized the long-context information.

This test specifically targets the model’s ability to retain and access information across very long sequences, providing a benchmark for evaluating its extended-context reasoning and retrieval performance.

It is important to note that strong performance in the "needle in a haystack" test does not guarantee retention of short-context performance across larger tasks. This evaluation is therefore limited in scope. We are actively working on developing more robust metrics and evaluation protocols to further enhance the model’s long-context capabilities.

---

Ethical Considerations and Limitations

The ALIA-40b-instruct model is an instruction-tuned variant with preliminary alignment. It has several limitations that users should be aware of. Ongoing work is addressing these areas, including comprehensive evaluation of societal and cognitive biases as well as safety, which will be reported in future updates.

Functional Limitations:

• No Function Calling: The model cannot natively execute or call external functions/APIs. Tasks requiring plugin calls or tool execution must be implemented outside the model.
• Reasoning & Math: The model is not guaranteed to perform robust chain-of-thought reasoning or advanced mathematics. Complex logical puzzles or multi-step inferences may fail or produce inconsistent answers.
• Code Generation: Although exposed to code during pretraining, ALIA-40b-Instruct is not a specialized code-generation model. It may produce code-like text, but outputs should be verified and tested before use in production codebases.
• Agentive Capabilities: The model does not have agentive or autonomous action capabilities. It cannot act as an autonomous agent or execute multi-step workflows.

Bias and Harm: The model may reflect biases present in its training data and may produce stereotyped, offensive, or otherwise harmful content, particularly regarding gender, ethnicity, religion, or other protected attributes. Work is ongoing to evaluate and mitigate societal and cognitive biases, and future releases will provide detailed reports on these analyses.

Safety and Alignment: The current alignment is preliminary and does not guarantee robust safety in all scenarios. The model may still follow malicious instructions or generate disallowed content if prompted. Additional filtering, human oversight, and alignment steps are essential. We are actively working to assess and improve the model’s safety, and a comprehensive report will be provided in subsequent updates.

Recommendations: Developers should implement additional safety filters, human oversight, targeted evaluation suites, and secondary evaluation models when deploying this model. Do not deploy ALIA-40b-Instruct in critical applications without extensive testing and mitigation. Users are responsible for assessing and mitigating harmful behavior or misinformation resulting from model outputs.

---

Additional information

Author

The Language Technologies Lab from Barcelona Supercomputing Center.

Contact

For further information, please send an email to [email protected] .

Copyright

Copyright(c) 2025 by Language Technologies Lab, Barcelona Supercomputing Center.

Funding

This work is funded by the Ministerio para la Transformación Digital y de la Función Pública - Funded by EU – NextGenerationEU within the framework of the project Modelos del Lenguaje.

This work has been promoted and supported by the Government of Catalonia through the Aina Project.

Acknowledgements

This project has benefited from the contributions of numerous teams and institutions, mainly through data contributions, knowledge transfer or technical support.

We are especially grateful to our ILENIA project partners: CENID, HiTZ and CiTIUS for their participation. We also extend our genuine gratitude to the Spanish Senate and Congress, Fundación Dialnet, and the ‘Instituto Universitario de Sistemas Inteligentes y Aplicaciones Numéricas en Ingeniería (SIANI)’ of the University of Las Palmas de Gran Canaria. Many other institutions have been involved in the project. Our thanks to Òmnium Cultural, Parlament de Catalunya, Institut d'Estudis Aranesos, Racó Català, Vilaweb, ACN, Nació Digital, El món and Aquí Berguedà. We thank the Welsh government, DFKI, Occiglot project, especially Malte Ostendorff, and The Common Crawl Foundation, especially Pedro Ortiz, for their collaboration.

We would also like to give special thanks to the NVIDIA team, with whom we have met regularly, especially to: Ignacio Sarasua, Adam Henryk Grzywaczewski, Oleg Sudakov, Sergio Perez, Miguel Martinez, Felipe Soares and Meriem Bendris. Their constant support has been especially appreciated throughout the entire process.

Their valuable efforts have been instrumental in the development of this work.

Disclaimer

Be aware that the model may contain biases or other unintended distortions. When third parties deploy systems or provide services based on this model, or use the model 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 Barcelona Supercomputing Center, as the owner and creator of the model, shall not be held liable for any outcomes resulting from third-party use.

Citation

@misc{gonzalezagirre2025salamandratechnicalreport,
      title={Salamandra Technical Report}, 
      author={Aitor Gonzalez-Agirre and Marc Pàmies and Joan Llop and Irene Baucells and Severino Da Dalt and Daniel Tamayo and José Javier Saiz and Ferran Espuña and Jaume Prats and Javier Aula-Blasco and Mario Mina and Adrián Rubio and Alexander Shvets and Anna Sallés and Iñaki Lacunza and Iñigo Pikabea and Jorge Palomar and Júlia Falcão and Lucía Tormo and Luis Vasquez-Reina and Montserrat Marimon and Valle Ruíz-Fernández and Marta Villegas},
      year={2025},
      eprint={2502.08489},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2502.08489}, 
}

License

Apache License, Version 2.0

Model Index