ALIA-40b-instruct-2606

[!NOTE] WARNING: Although this model has undergone safety and value alignment, it may still occasionally generate unintended or undesired outputs.

Sampling Parameters: For optimal performance, we recommend using temperatures close to zero (0 - 0.2). Additionally, we advise against using any type of repetition penalty, as from our experience, it negatively impacts instructed model's responses.

ALIA-40b-instruct Model Card

ALIA-40b-instruct-2606 is the latest release in the ALIA model family.

Main improvements

• Instruction Following: Enhanced alignment and instruction-tuning, leading to more reliable adherence to user intent across a wide range of tasks.
• Input Robustness: Strengthened resilience to noisy, ambiguous, or malformed user inputs, resulting in more stable and predictable responses.
• Safety: Improved safety alignment, reducing the likelihood of generating responses related to sensitive or restricted topics and improving resistance to attacks, while maintaining helpfulness on allowed content.
• Multi-turn: The model is more robust and contextually aware throughout extended conversations.
• System prompt following: Enhanced adherence allows the model to better maintain constraints and personas.

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.

The model has been made openly available under the Apache 2.0 license for the benefit of society. Through this open release, ALIA seeks to accelerate scientific progress and innovation by enabling the scientific community and society at large to use, study, evaluate, adapt, further develop, and build upon the model. This open approach promotes collaboration, transparency, and the responsible development of artificial intelligence, helping to advance knowledge, drive technological innovation, and contribute to the public interest. The post-training datasets are publicly available in ALIA-SFT, ALIA-SFT-tools, ALIA-DPO-helpulness, and ALIA-DPO-safety.

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 development 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) an alignment stage to better match human preferences and safety.

After the long-context adaptation, the post-training process begins with a supervised fine-tuning (SFT) stage, performed over 714k conversation samples to strengthen instruction following and add conversational capabilities.

In the third stage, the model is aligned with human preferences through Direct Policy Optimization (DPO) using a mixture of 73k preference pairs. Of this mixture, approximately 59% of the pairs target general model helpfulness, 24% focus on response safety and 16% focus on Exact Instruction Following. This represents an over 80% reduction compared to the previous mixture, resulting in a smaller but substantially higher-quality set of preference pairs.

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

Supervised Fine-Tuning (SFT)

Alignment (DPO)

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. ALIA-40b-instruct is best suited for tasks like multilingual chatbots, summarization, translation, and content generation, provided users are aware of its limitations.

The ALIA‑40b‑instruct model has been developed in a research and development environment as part of an experimental scientific project by BSC-CNS, a public research organization. As such, ALIA-40b-instruct is provided strictly on an “as-is” basis, without warranties of any kind, express or implied. This includes, without limitation, any warranties as to accuracy, quality, completeness, reliability, availability, non-infringement, merchantability, fitness for a particular purpose, suitability for commercial deployment, production use, or regulatory compliance.

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 two complementary frameworks, each selected to best support its corresponding stage:

• Supervised Fine-Tuning (SFT): Conducted in NeMo-RL, with minor modifications to adapt it to our infraestructure.
• 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 model can be used either directly in Python using the transformers library or deployed as a service and used through standard API calls.

While the former gives the most control over the inference process it requires the code to be executed on a machine with a sufficiently powerful GPU to run the model locally, and is more error prone than the alternative. We therefore strongly recommend the latter, as deploying the model as a service can be done either locally or on a remote server and makes the model available to multiple clients in parallel among other advantages.

Unless you have very specific needs (e.g. for research) that require adapting the inference process it is preferable to follow the "deployment as a service" guidelines below.

Local inference with Python / transformers

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 torch

model_id = "BSC-LT/ALIA-40b-instruct-2606"

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 system for instructions that guide the model's behavior, user, for content supplied by the user, or assistant for the model's responses), and finished with the <|im_end|> token:

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

Loading the model with transformers' AutoModelForCausalLM guarantees that adequate sampling parameters are used during generation. If using alternative inference libraries such as vLLM, Ollama, or SGLang, it is crucial to verify that optimal parameters are used. To this end, in order to ensure optimal results, we recommend using temperatures around 0-0.2 without any type of repetition penalties applied.

---

Deployment as service and remote use (Messages API)

In our experience, vllm works well for deploying the full unquantized version of the model, whereas llama.cpp is appropriate for the quantized (GGUF) version. We strongly discourage using ollama as we have encountered compatibility issues that may seriously degrade the model's performance.

The easiest and most reliable way to have a working deployment of ALIA-40b-instruct is through the "Deploy / HF Inference Endpoints" option directly on the Hugging Face model page. This automatically creates a functioning endpoint, using vllm or llama.cpp according to the model variant, with an appropriately dimensioned GPU. While there are additional settings available for the endpoint we found the standard configuration proposed by Hugging Face to be a reasonable starting point.

Once the endpoint is running, the model can be easily called using OpenAI's "Messages API" (the de facto standard API for LLM use). By using this API the chat template is applied automatically by the service, requiring no explicit configuration on the client side. The endpoint's configuration page on Hugging Face also provides a "Playground" for testing and API examples, as well as a simple chat interface.

Example usage:

# pip install openai 

from openai import OpenAI 

client = OpenAI(
	base_url = YOUR_ENDPOINT_URL,
	api_key = YOUR_HF_TOKEN
)

chat_completion = client.chat.completions.create(
	model = "BSC-LT/ALIA-40b-instruct-2606",
	messages = [
		{
			"role": "user",
			"content": "What is deep learning?"
		}
	],
	stream = True,
    max_tokens = 1000,
    temperature=0.1
)

print(chat_completion.choices[0].message.content)

The model can also be deployed locally or on any server infrastructure with sufficient GPUs, using vllm or llama.cpp. We recommend an initial deployment on Hugging Face as a point of reference and comparison to make sure the model is behaving as expected in the desired deployment setup.

[!NOTE] To check that your endpoint is working correctly, you can try to replicate the examples contained in this Colab Notebook

---

Instruction Tuning Data

The dataset used in the supervised fine-tuning stage consists of 714k conversations. 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 language and contribution to the overall corpus:

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:

<details> <summary>SFT Datasets</summary> <table> <tr> <th>Dataset</th> <th>Generation Method</th> <th>License</th> <th>Description</th> </tr> <tr> <td>aya-dataset</td> <td>Human Crowdsourced</td> <td>Apache-2.0</td> <td><a href="https://huggingface.co/datasets/CohereLabs/aya_dataset">aya_dataset</a> for the languages of interest.*</td> </tr> <tr> <td>coqcat-train</td> <td>Human Annotation</td> <td>CC-BY-NC-ND-4.0</td> <td><a href="https://huggingface.co/datasets/projecte-aina/CoQCat">CoQCat</a> train split, formatted using conversational templates.</td> </tr> <tr> <td>databricks-dolly-15k</td> <td>Human Annotation</td> <td>CC-BY-SA-3.0</td> <td><a href="https://huggingface.co/datasets/databricks/databricks-dolly-15k">databricks-dolly-15k</a> dataset.*</td> </tr> <tr> <td>dolly-ca</td> <td>Human Translation</td> <td>CC-BY-SA-3.0</td> <td><a href="https://huggingface.co/datasets/projecte-aina/dolly3k_ca">dolly3k_ca</a> dataset.</td> </tr> <tr> <td>flores-dev</td> <td>Human</td> <td>CC-BY-SA-4.0</td> <td>Flores-200 dev split, formatted using conversational templates.</td> </tr> <tr> <td>mentor-es</td> <td>Human Annotation</td> <td>CC-BY-SA-4.0</td> <td><a href="https://huggingface.co/datasets/projecte-aina/MentorES">MentorES</a> dataset.</td> </tr> <tr> <td>mentor-ca</td> <td>Machine Translation</td> <td>CC-BY-SA-4.0</td> <td><a href="https://huggingface.co/datasets/projecte-aina/MentorCA">MentorCA</a> dataset. Machine translated version of MentorES.</td> </tr> <tr> <td>no-robots</td> <td>Human Annotation</td> <td>CC-BY-NC-4.0</td> <td><a href="https://huggingface.co/datasets/HuggingFaceH4/no_robots">no_robots</a> dataset.*</td> </tr> <tr> <td>rag-multilingual</td> <td>Synthetic</td> <td>CC-BY-SA-4.0</td> <td><a href="https://huggingface.co/datasets/projecte-aina/RAG_Multilingual">RAG_Multilingual</a> dataset. Synthetic QA dataset generated with Mixtral8x7b.</td> </tr> <tr> <td>tower-blocks</td> <td>Mixture</td> <td>Various licenses (only open licensed instances are used)</td> <td><a href="https://huggingface.co/datasets/Unbabel/TowerBlocks-v0.2">TowerBlocks-v0.2</a> filtered by subdataset license and the languages of interest.*</td> </tr> <tr> <td>oasst2_self-identity-rephrase</td> <td>Human Crowdsourced / Synthetic</td> <td>Apache-2.0</td> <td>Identity instances from <a href="https://huggingface.co/datasets/OpenAssistant/oasst2">oasst2</a> dataset for the languages of interest. Subsequently rephrased to adapt the model’s identity information to our case using DeepSeek-V3-0324.</td> </tr> <tr> <td>self-identity</td> <td>Synthetic</td> <td>Apache-2.0 (internal)</td> <td>Conversations involving self-identity information of the model, synthetically curated using DeepSeek-V3-0324.</td> </tr> <tr> <td>open-r1-math</td> <td>Synthetic</td> <td>Apache-2.0</td> <td>Default 93k split of the <a href="https://huggingface.co/datasets/open-r1/OpenR1-Math-220k">OpenR1-Math-220k</a> dataset.*</td> </tr> <tr> <td>open-r1-math_translated</td> <td>Synthetic</td> <td>Apache-2.0 (internal)</td> <td>OpenR1-Math-220k default split translated to the languages of interest with DeepSeek-V3-0324.</td> </tr> <tr> <td>fineweb-edu_qa</td> <td>Synthetic</td> <td>Apache-2.0 (internal)</td> <td>QA conversations created by prompting DeepSeek-V3-0324 with the highest quality documents of <a href="https://huggingface.co/datasets/HuggingFaceFW/fineweb-edu">FineWeb-Edu</a>. Subsequently filtered with the same model to ensure self-contained question-answering pairs meet quality thresholds.</td> </tr> <tr> <td>wildchat-curated-deepseekv3</td> <td>Human / Synthetic</td> <td>Apache-2.0 (internal)</td> <td>Human prompts from the <a href="https://huggingface.co/datasets/allenai/WildChat-1M">WildChat-1M</a> dataset together with responses generated with DeepSeek-V3-0324.</td> </tr> <tr> <td>eif-augmentation</td> <td>Synthetic</td> <td>Apache-2.0 (internal)</td> <td>Conversations sampled from the mixture and modified to target exact instruction following. Generated with DeepSeek-V3-0324 and filtered with verifiable constrains.</td> </tr> <tr> <td>multiturn-augmentation</td> <td>Synthetic</td> <td>Apache-2.0 (internal)</td> <td>Conversations sampled from the mixture and augmented with subsequent user/assistant turns with DeepSeek-V3-0324.</td> </tr> <tr> <td>long-context-qa</td> <td>Synthetic</td> <td>Apache-2.0 (internal)</td> <td>QA conversations created by prompting previous internal model versions with long documents extracted from <a href="https://huggingface.co/datasets/HuggingFaceFW/fineweb-edu">FineWeb-Edu</a>.</td> </tr> <tr> <td>system-prompt-eif-translation</td> <td>Human</td> <td>CC-BY-4.0 (internal)</td> <td>Long-context translation conversations specially formatted to target system-prompt following and special formatting capabilities. Original translations extracted from <a href="https://huggingface.co/datasets/BSC-LT/ACAData">ACAData</a> dataset.</td> </tr> <tr> <td>system-prompt_multi-turn_non-verifiable</td> <td>Human</td> <td>Apache-2.0 (internal)</td> <td>Initial prompts sampled from the remaining training mixture and augmented with manually written system prompts. Subsequently expanded with synthetic user and assistant turns using DeepSeek-V3-0324.</td> </tr> </table> <p>*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.</p> </details>

Alignment Data

The preference data for alignment was synthetically generated, resulting in a multilingual corpus of approximately 70k prompts designed to improve both helpfulness and safety.

• Helpfulness Prompts: Prompts include generic instruction following, exact instruction following, mathematics, question answering, and reasoning tasks across Catalan, Spanish, English, Euskera, and Galician. These prompts were sampled from our SFT mix.
• Safety Sprompts: Prompts were synthetically generated from seed prompts written by human annotators, covering 12 categories from the AI Luminate Hazard Taxonomy, to ensure broad coverage of safety-related scenarios. These attack prompts were generated in Catalan, Spanish, English, Euskera, and Galician, using different Qwen3 Models and Deepseek-V3-0324.

Following approaches similar to Tulu3/UltraFeedback and PKU-Safe-RLHF, each prompt 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.

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 12 harm categories:

Disaggregated by harm category:

Model Pool for Synthetic Data Generation

The model pool used for synthetic data generation consists of previous public releases, such as ALIA-40b-instruct-2605, together with a selection of internal ALIA model checkpoints. Including models from different stages of the ALIA development cycle increases response diversity and reduces the likelihood of overfitting the synthetic preference data to the behavior of a single model family.

Evaluation

This section presents the main capabilities evaluated in ALIA-40B-Instruct, the LLM-as-a-Judge setup, and the results obtained across the three evaluation protocols: gold-standard benchmarking, LLM-as-a-Judge evaluation, and human evaluation.

Description of evaluated capacities

The model is evaluated across a set of predefined capabilities using three complementary approaches: gold-standard benchmarking, LLM-as-a-Judge evaluation, and human evaluation. The evaluation setup and datasets used for each capability are described below.

• C1 – Instruction Following: Evaluates the model’s ability to follow complex natural-language instructions through LLM-as-a-Judge and human evaluation.

  • LLM-as-a-Judge: 270 instances sampled from the complex_if_single_turn_v5 subset of AdvancedIF, excluding format-related instances.
  • Human evaluation: 20 instances sampled from InFoBench and AdvancedIF. The InFoBench instances are balanced across the Easy and Hard subsets, while the AdvancedIF instances cover the complex_if_single_turn_v5 subset.

• C2 – Format Following: Evaluates the model’s ability to comply with structured-output requirements through gold-standard benchmarking and LLM-as-a-Judge evaluation.

  • Dataset: 200 instances sampled from FoFo, covering JSON, CSV, XML, and YAML, together with a small number of JSON tasks from MultiIF. The same dataset is used for both evaluation methods.

• C3 – Multi-Turn Context Management: Evaluates the model’s ability to maintain and correctly use conversational context through gold-standard benchmarking, LLM-as-a-Judge, and human evaluation.

  • Benchmarking and LLM-as-a-Judge: 70 instances sampled from AdvancedIF and balanced across five task types: Anaphora Resolution, Self-Correction, Self-Affirmation, Instruction Clarification, and Content Confusion.
  • Human evaluation: 20 instances sampled from MTBench101 and balanced across the same task types as the automatic evaluation dataset.

• C5 – Function Calling: Evaluates the model’s ability to determine whether a user request requires an external tool call and to generate the correct call or provide an appropriate alternative response when no tool should be used. This capability is evaluated through gold-standard benchmarking and LLM-as-a-Judge.

  • Dataset: 200 instances sampled from When2Call and balanced across the target labels cannot_answer, request_for_information, and tool_call.

LLM-as-a-Judge setup

We use gpt-oss-120b 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. 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 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."""
}

Results

<table> <thead> <tr> <th>Category</th> <th>Task</th> <th>Benchmarking Results</th> <th>LLM-as-a-Judge Results</th> <th>Human Evaluation Results</th> </tr> </thead> <tbody> <tr> <td><strong>C1 – Instruction Following</strong></td> <td><strong>AdvancedIF</strong></td> <td align="center">x</td> <td align="center">69.22</td> <td align="center">86.0</td> </tr> <tr> <td rowspan="4"><strong>C2 – Format Following</strong></td> <td><strong>CSV</strong></td> <td align="center">59.57</td> <td align="center">6.38</td> <td align="center">x</td> </tr> <tr> <td><strong>XML</strong></td> <td align="center">25.53</td> <td align="center">4.26</td> <td align="center">x</td> </tr> <tr> <td><strong>JSON</strong></td> <td align="center">62.29</td> <td align="center">18.03</td> <td align="center">x</td> </tr> <tr> <td><strong>YAML</strong></td> <td align="center">77.77</td> <td align="center">15.56</td> <td align="center">x</td> </tr> <tr> <td rowspan="2"><strong>C3 – Multi-Turn Context Management</strong></td> <td><strong>AdvancedIF</strong></td> <td align="center">x</td> <td align="center">59.33</td> <td rowspan="2" align="center">97.0</td> </tr> <tr> <td><strong>MT-Bench-101</strong></td> <td align="center">x</td> <td align="center">79.50</td> </tr> <tr> <td><strong>C5 – Function Calling</strong></td> <td><strong>When2Call</strong></td> <td align="center">79.68</td> <td align="center">20.50</td> <td align="center">x</td> </tr> </tbody> </table>

Results are reported on a 0–100 scale, where higher values indicate better performance. For the LLM-as-a-Judge evaluation, gpt-oss-120b was used as the judge model in all cases. Cells marked with x indicate capabilities that were not evaluated using the method specified in the corresponding column.

---

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.

Functional Limitations:

• No Function Calling: The model cannot natively execute or call external functions/APIs. Refer to the function-calling version for these use cases: ALIA-40b-fc-2606
• 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.

Safety and Alignment:

Alignment has been substantially enhanced compared to earlier versions. However, no large language model can be guaranteed to behave safely in every possible context or adversarial setting. As a result, the model may still exhibit unsafe behavior in certain edge cases, including responding to malicious prompts or generating disallowed content. To evaluate the model's vulnerabilities, we conduct a Red Teaming assessment using adversarial prompts datasets written by our annotation team, and with DeepseekV3-0324 serving as the moderator model (LLM-as-a-judge with a judge prompt also validated by our annotation team). This evaluation is carried out in Spanish, Catalan, English, Basque, and Galician. Results yielded an average attack success rate of 4.5%.

Additional filtering, human oversight, and alignment steps are essential. We are actively working to improve and assess the model's safety, including human annotation and evaluation, as well as the development of multilingual safety datasets. 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, and ensuring compliance with applicable regulations, including those governing the use of Artificial Intelligence.

---

Additional information

Author

The Language Modeling team from AI Institute at Barcelona Supercomputing Center.

Contact

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

Copyright

Copyright(c) 2026 by The Language Modeling team from AI Institute at 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: Marcelo Sanchez, 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