Carvalho_pt-gl-1.3B

Carvalho_pt-gl-1.3B is a 1.3B-parameter transformer-based causal language model for Galician and European Portuguese. It is the result of a continual pretraining of a Cerebras-GPT-1.3B adapted to catalan, spanish and english previously by the AINA Project.

This model is part of the Carvalho familily, a family of LLMs specialized in Portuguese and Galician which can be found here.

How to use

import torch
from transformers import pipeline, AutoTokenizer, AutoModelForCausalLM

token_HF=""#Obter na páxina de HuggingFace
input_text = "Hoxe fai un bo día. O sol  "

model_id  = "Nos-PT/Carvalho_pt-gl-1.3B"
tokenizer = AutoTokenizer.from_pretrained(model_id, use_auth_token=token_HF)
model = AutoModelForCausalLM.from_pretrained(model_id, use_auth_token=token_HF)
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 5B tokens of texts in Galician and European Portuguese, whose main sources are CorpusNÓS for Galician and Arquivo.pt for Portuguese.

Training hyperparameters

• seed: 42
• 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: 5e-05
• num_epochs: 1.0

Framework

The training was conducted on the Galician Supercomputing Center (CESGA), using various nodes with 2 GPUs NVIDIA A100 40G.

Evaluation

The evaluation was carried out using the Calame dataset, in its Portuguese and Galician versions, and in the GLUE benchmark. In the latter case, the evaluated models were fine-tuned to rotate the tasks.

Models	CALAME-GL	CALAME-PT
Carvalho_pt-gl 1.3B	0.397	0.455
GlorIA 1.3B	0.219	0.488
Gervasio-PTPT 7B	0.265	0.434
mGPT 1.3B	0.264	0.425
Bloom-1b1	0.262	0.456
Cerebras-GPT 1.3B	0.177	0.167

Models	RTE	MRPC	STS-B	WNLI
	Acc	F1	Pearson	Acc
Encoder-only
AiBERTa Base	55.3	83.2	80.2	58.9
Albertina-PTPT 100m	55.4	87.6	84.5	65.1
Albertina-PTPT 900m	80.6	89.8	88.7	65.1
Albertina-PTPT 1.5B	82.9	90.3	88.7	59.6
Decoder-only
Carvalho_pt-gl 1.3B	68.0	86.0	82.6	65.1
Gloria 1.3B	63.8	85.2	82.0	65.1
Gervásio 7B	83.2	90.5	87.9	64.4
Bloom 1.1B	71.5	87.7	85.7	63.6
mGPT 1.3B	58.9	85.5	78.3	65.1

Additional information

Contact

For further information, please send an email to

License

MIT License

Copyright (c) 2024

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

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.

Cite this model

@inproceedings{gamallo2024galician,
  title={A Galician-Portuguese Generative Model},
  author={Gamallo, Pablo and Rodr{\'\i}guez, Pablo and Santos, Daniel and Sotelo, Susana and Miquelina, Nuno and Paniagua, Silvia and Schmidt, Daniela and de-Dios-Flores, Iria and Quaresma, Paulo and Bardanca, Daniel and others},
  booktitle={EPIA Conference on Artificial Intelligence},
  pages={292--304},
  year={2024},
  organization={Springer}
}