Llama-3.1-Carballo

Table of Contents

• Llama-3.1-Carballo
  • Table of Contents
  • Model description
  • Intended uses and limitations
  • How to use
  • Training
    • Tools
    • Training data
    • Training hyperparameters
    • Framework
  • Evaluation
  • Additional information
    • Contact
    • License
    • Funding

Model description

Llama-3.1-Carballo is a 8B-parameter transformer-based causal language model for Galician, Portuguese, Spanish, Catalan and English. It is the result of a continual pretraining of meta-llama/Llama-3.1-8B with a multilingual corpus of almost 20B tokens, with an emphasis on Galician texts.

This model is part of the Carballo familily, a family of LLMs specialized in Galician. Smaller models can be found here

Intended uses and limitations

The Llama-3.1-Carballo model is ready-to-use only for causal language modeling. It can perform text-generation tasks and be fine-tuned for specific scenarios.

How to use

import torch
from transformers import pipeline, AutoTokenizer, AutoModelForCausalLM

input_text = "Hoxe fai un bo día. O sol  "

model_id  = "proxectonos/Llama-3.1-Carballo"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)
generator = pipeline(
    "text-generation",
    model=model,
    tokenizer=tokenizer,
    torch_dtype=torch.bfloat16,
    trust_remote_code=True,
    device_map="auto",
)
generation = generator(
    input_text,
    do_sample=True,
    top_k=10,
    eos_token_id=tokenizer.eos_token_id
)

print(f"Result: {generation[0]['generated_text']}")

Training

Tools

It was trained using HuggingFace Transformers and Pytorch, using the Causal Modeling Language script. We also use DeepSpeed to deal with the huge size of the model.

Training data

The training corpus consists of texts in 5 languages, with an emphasis on Galician. The main aim of this is to ensure that the model learns to work with this language perfectly, while maintaining knowledge of languages already known (Spanish, English), learning others (Catalan) or adapting existing language varieties (Portuguese-PT instead of Portuguese-BR).

The corpus is structured as follows:

Training hyperparameters

• seed: 42
• num_devices: 5
• train_batch_size: 4
• eval_batch_size: 4
• gradient_acummulation: 8
• optimizer: AdamW
• betas: (0.9,0.999)
• epsilon: 1e-08
• weight_decay_rate: 0.1
• scheduler: "Linear"
• learning_rate: 1e-04
• num_epochs: 1.0

Framework

The training was conducted in the Galicia Supercomputing Center (CESGA), using 5 nodes with 2 GPUs NVIDIA A100 each one.

Evaluation

In process...

Additional information

Contact

For further information, please send an email to proxecto.nos@usc.gal

Funding

This model was development within the Nós Project, funded by the Ministerio para la Transformación Digital y de la Función Pública - Funded by EU – NextGenerationEU within the framework of the project ILENIA with reference 2022/TL22/00215336.

Citation

@inproceedings{rodriguez-etal-2025-continued,
    title = "Continued Pretraining and Interpretability-Based Evaluation for Low-Resource Languages: A {G}alician Case Study",
    author = "Rodr{\'i}guez, Pablo  and
      Su{\'a}rez, Silvia Paniagua  and
      Gamallo, Pablo  and
      Docio, Susana Sotelo",
    editor = "Che, Wanxiang  and
      Nabende, Joyce  and
      Shutova, Ekaterina  and
      Pilehvar, Mohammad Taher",
    booktitle = "Findings of the Association for Computational Linguistics: ACL 2025",
    month = jul,
    year = "2025",
    address = "Vienna, Austria",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2025.findings-acl.240/",
    doi = "10.18653/v1/2025.findings-acl.240",
    pages = "4622--4637",
    ISBN = "979-8-89176-256-5",
    abstract = "Recent advances in Large Language Models (LLMs) have led to remarkable improvements in language understanding and text generation. However, challenges remain in enhancing their performance for underrepresented languages, ensuring continual learning without catastrophic forgetting, and developing robust evaluation methodologies. This work addresses these issues by investigating the impact of Continued Pretraining (CPT) on multilingual models and proposing a comprehensive evaluation framework for LLMs, focusing on the case of Galician language. Our first contribution explores CPT strategies for languages with limited representation in multilingual models. We analyze how CPT with Galician corpora improves text generation while assessing the trade-offs between linguistic enrichment and task-solving capabilities. Our findings show that CPT with small, high-quality corpora and diverse instructions enhances both task performance and linguistic quality. Our second contribution is a structured evaluation framework based on distinguishing task-based and language-based assessments, leveraging existing and newly developed benchmarks for Galician. Additionally, we contribute new Galician LLMs, datasets for evaluation and instructions, and an evaluation framework."
}