Llama-3.1-8B-Instruct-Galician a.k.a. Cabuxa 2.0

This model is a continued pretraining version of meta-llama/Llama-3.1-8B-Instruct on the CorpusNós dataset.

Model Description

• Developed by: UDC Information Retrieval Lab (IRLab)
• Language(s) (NLP): Multilingual, adapted to Galician
• License: llama3.1
• Finetuned from model: meta-llama/Llama-3.1-8B-Instruct
• Repository: Adapting Large Language Models for Underrepresented Languages
• Paper: Coming soon

How to Get Started with the Model

import transformers
import torch

model_id = "irlab-udc/Llama-3.1-8B-Instruct-Galician"

pipeline = transformers.pipeline(
    "text-generation",
    model=model_id,
    model_kwargs={"torch_dtype": torch.bfloat16},
    device_map="auto",
)

messages = [
    {"role": "system", "content": "You are a conversational AI that always responds in Galician."},
    {"role": "user", "content": "Cal é a principal vantaxe de usar Scrum?"},
]

outputs = pipeline(messages, max_new_tokens=512)

print(outputs[0]["generated_text"][-1]["content"])

Training Hyperparameters

Parameter	Value
learning_rate	0.0001
train_batch_size	32
eval_batch_size	1
seed	42
distributed_type	multi-GPU
num_devices	4
gradient_accumulation_steps	2
total_train_batch_size	256
total_eval_batch_size	4
optimizer	Adam with betas=(0.9, 0.999), epsilon=1e-08
lr_scheduler_type	cosine
lr_scheduler_warmup_ratio	0.1
num_epochs	1.0

Training results

Training Loss	Epoch	Step	Validation Loss
2.0606	0.1682	900	2.0613
1.9898	0.3363	1800	1.9929
1.9847	0.5045	2700	1.9613
1.9577	0.6726	3600	1.9445
1.9287	0.8408	4500	1.9368

Environmental Impact

Carbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).

• Hardware Type: 4x NVIDIA A100 SXM4 80 GB (TDP of 400W)
• Hours used: 60
• Cloud Provider: Private infrastructure
• Carbon Emitted: 10.37 Kg. CO₂ eq.

Citation

@inproceedings{bao-perez-parapar-xovetic-2024,
  title={Adapting Large Language Models for Underrepresented Languages},
  author={Eliseo Bao and Anxo Pérez and Javier Parapar	},
  booktitle={VII Congreso XoveTIC: impulsando el talento cient{\'\i}fico},
  year={2024},
  organization={Universidade da Coru{\~n}a, Servizo de Publicaci{\'o}ns}
  abstact = {The popularization of Large Language Models (LLMs), especially with the development of conversational systems, makes mandatory to think about facilitating the use of artificial intelligence (AI) to everyone. Most models neglect minority languages, prioritizing widely spoken ones. This exacerbates their underrepresentation in the digital world and negatively affects their speakers. We present two resources aimed at improving natural language processing (NLP) for Galician: (i) a Llama 3.1 instruct model adapted through continuous pre-training on the CorpusNos dataset; and (ii) a Galician version of the Alpaca dataset, used to assess the improvement over the base model. In this evaluation, our model outperformed both the base model and another Galician model in quantitative and qualitative terms}
}