You can deploy this model and start earning money today!

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

Work In Progress New versions will be available during the coming weeks/months.

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-2601 is the latest release in the ALIA model family. While development is ongoing and further updates are expected, this version already incorporates several notable improvements over previous releases.

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.
Long-Context: Improved long-context capabilities compared to previous version ALIA-40b-instruct-2512

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 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 808k 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 368k preference pairs. Of this mixture, approximately 82% of the pairs target general model helpfulness, while 18% focus on response safety.

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

Hyperparameter	Value
Learning rate	9e-7
LR Scheduler	Constant
Tokens per update	4M
Training tokens (4k →32k).	2B
Training tokens (32k →160k).	36.8B

Supervised Fine-Tuning (SFT)

Hyperparameter	Value
Learning rate	5e-6
Batch size	1024
Epochs	1
LR Scheduler	Cosine
Warmup Ratio	0.03
NEFTune Noise Alpha	5
Number of Samples	807,750

Alignment

Hyperparameter	Value
Learning rate	2e-6
Batch size	1024
Epochs	2
Beta	0.1
LR Scheduler	Linear
Number of samples	368,475

Architecture

Attribute	Value
Total Parameters	40,433,885,184
Embedding Parameters	2,097,152,000
Layers	48
Hidden size	8,192
Attention heads	64
Context length	163,840
Vocabulary size	256,000
Precision	bfloat16
Embedding type	RoPE
Activation Function	SwiGLU
Layer normalization	RMS Norm
Flash attention	✅
Grouped Query Attention	✅
Num. query groups	8

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 two 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.
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:

Phase	Nodes	GPUs
SFT	16	64
Alignment	16	64

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:

python

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_id = "BSC-LT/ALIA-40b-instruct-2601"

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|>

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:

python

# 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-2601",
	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 808k 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:

Dataset	ca	en	es	eu	gl	pt	Total Conversations
aya-dataset		3940	3851	939		8995	17725
coqcat-train	4797						4797
databricks-dolly-15k		15007					15007
dolly-ca	3232						3232
flores-dev	986	1037	1964	493	505		4985
mentor-ca	7119						7119
mentor-es			7122				7122
no-robots		9477					9477
rag-multilingual	16043	14996	11263				42302
tower-blocks		7762	1000			1000	9762
oasst2_self-identity-rephrase	7	1074	447	8			1536
self-identity	1900	1978	1943	1927	1880		9628
open-r1-math		92960					92960
open-r1-math_translated	46357		92601	46361	46431	46434	278184
fineweb-edu_qa	23374	20803	23311	22283	22307		112078
wildchat-curated-deepseekv3		173948	17888				191836
Total	103815	342982	161390	72011	71123	56429	807750

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	Identity instances from 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.
wildchat-curated-deepseekv3	Human / Synthetic	Apache-2.0 (internal)	Human prompts from the WildChat-1M dataset together with responses generated with DeepSeek-V3-0324.

*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:

Multiple responses were produced using a pool of permissively licensed models (see Model Pool) on helpfulness or safety, depending on the prompt.
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.
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:

Dataset	ca	en	es	eu	gl	Total
aya	0	2 586	3 019	902	0	6 507
coqcat	4 448	0	0	0	0	4 448
dolly	0	9 925	0	0	0	9 925
dolly-ca	2 971	0	0	0	0	2 971
flores-dev	1 219	589	1 786	357	457	4 408
identity	2 924	20 120	15 720	2 396	2 276	43 436
m-personas	2 674	1 215	2 852	2 791	2 530	12 062
mentor-ca	6 517	0	0	0	0	6 517
mentor-es	0	0	6 007	0	0	6 007
new_open-orca	0	15 528	0	0	0	15 528
no-robots-system-prompt	0	5 913	0	0	0	5 913
oasst-ca	2 195	0	0	0	0	2 195
persona-generic	8 849	0	9 464	8 899	8 588	35 800
persona-reasoning	8 721	0	9 501	8 977	8 474	35 673
rag-multilingual	15 072	10 003	9 955	0	0	35 030
tower-blocks	0	4 126	692	0	0	4 818
Total	55 590	170 005	58 996	24 322	22 325	231 238

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

Language	Instances
ca	21074
es	20887
en	6370
eu	13459
gl	9951

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

task	metric	result
belebele_spa_Latn	acc	0.72
cocoteros_es	bleu	0.05
cocoteros_es	rouge1	0.32
copa_es	acc	0.75
escola	mcc	0
flores_es	bleu	0.25
mgsm_direct_es_spanish_bench	exact_match,flexible-extract	0.08
mmmlu_es	acc	0.41
openbookqa_es	acc	0.24
paws_es_spanish_bench	acc	0.32
phrases_es-va	bleu	0.59
phrases_va-es	bleu	0.69
wnli_es	acc	0.01
xlsum_es	bleu	0.02
xnli_es_spanish_bench	acc	0.28
xquad_es	f1	0.07
xstorycloze_es	acc	0.52

Catalan

task	metric	result
arc_ca_challenge	acc	0.37
arc_ca_easy	acc	0.71
belebele_cat_Latn	acc	0.71
cabreu_abstractive	bleu	0.04
cabreu_extractive	rouge1	0.15
cabreu_extreme	bleu	0.03
catalanqa	f1	0.11
catcola	mcc	0
cocoteros_va	bleu	0.05
cocoteros_va	rouge1	0.33
copa_ca	acc	0.74
coqcat	f1	0.6
flores_ca	bleu	0.31
mgsm_direct_ca	exact_match,flexible-extract	0.28
openbookqa_ca	acc	0.2
parafraseja	acc	0.36
paws_ca	acc	0.41
phrases_ca-va	bleu	0.75
phrases_va-ca	bleu	0.83
piqa_ca	acc	0.53
siqa_ca	acc	0.3
teca	acc	0.39
wnli_ca	acc	0.13
xquad_ca	f1	0.11
xstorycloze_ca	acc	0.52

Basque

task	metric	result
arc_eu_challenge	acc	0.27
arc_eu_easy	acc	0.55
belebele_eus_Latn	acc	0.67
eus_proficiency	acc	0.33
eus_reading	acc	0.5
eus_trivia	acc	0.5
flores_eu	bleu	0.19
mgsm_direct_eu	exact_match,flexible-extract	0.16
mgsm_native_cot_eu	exact_match,get-answer	0
paws_eu	acc	0.34
piqa_eu	acc	0.39
qnlieu	acc	0.19
wnli_eu	acc	-0.1
xcopa_eu	acc	0.51
xnli_eu	acc	0.3
xnli_eu_native	acc	0.3
xstorycloze_eu	acc	0.38

Galician

task	metric	result
belebele_glg_Latn	acc	0.73
flores_gl	bleu	0.29
galcola	mcc	0
mgsm_direct_gl	exact_match,flexible-extract	0.1
openbookqa_gl	acc	0.16
parafrases_gl	acc	0.22
paws_gl	acc	0.4
summarization_gl	bleu	0.04
xnli_gl	acc	0.37
xstorycloze_gl	acc	0.49

English

task	metric	result
arc_challenge	acc	0.4
arc_easy	acc	0.73
belebele_eng_Latn	acc	0.77
cola	mcc	0
copa	acc	0.78
hellaswag	acc	0.54
hellaswag	acc_norm	-0.32
mgsm_direct_en	exact_match,flexible-extract	0.09
mmlu	acc	0.45
openbookqa	acc	0.18
paws_en	acc	0.44
piqa	acc	0.65
social_iqa	acc	0.25
truthfulqa_mc1	acc	0.19
truthfulqa_mc2	acc	0.41
wnli	acc	0.41
xnli_en_iberobench	acc	0.36
xquad_en	f1	0.17
xstorycloze_en	acc	0.6

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:

python

"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:

python

"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):

python

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

WiP

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.

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:

WiP

Safety and Alignment:

Alignment has been substantially enhanced compared to earlier versions, though it is not yet complete. 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 13.3%.

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 Technologies Lab from Barcelona Supercomputing Center.

Contact

For further information, please send an email to langtech@bsc.es.

Copyright

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

Model	Base	Instruct
2b	Link	Link
7b	Link	Link
40b	Link	Link