language:
- pt
license: apache-2.0
library_name: transformers
tags:
- text-generation-inference
datasets:
- TucanoBR/GigaVerbo
metrics:
- perplexity
pipeline_tag: text-generation
widget:
- text: A floresta da Amazônia é conhecida por sua
example_title: Exemplo
- text: Uma das coisas que Portugal, Angola, Brasil e Moçambique tem em comum é o
example_title: Exemplo
- text: O Carnaval do Rio de Janeiro é
example_title: Exemplo
inference:
parameters:
repetition_penalty: 1.2
temperature: 0.2
top_k: 20
top_p: 0.2
max_new_tokens: 150
co2_eq_emissions:
emissions: 350000
source: CodeCarbon
training_type: pre-training
geographical_location: Germany
hardware_used: NVIDIA A100-SXM4-80GB
model-index:
- name: Tucano-630m
results:
- task:
type: text-generation
name: Text Generation
dataset:
name: CALAME-PT
type: NOVA-vision-language/calame-pt
split: all
args:
num_few_shot: 0
metrics:
- type: acc
value: 56.55
name: accuracy
source:
url: https://huggingface.co/datasets/NOVA-vision-language/calame-pt
name: Context-Aware LAnguage Modeling Evaluation for Portuguese
- task:
type: text-generation
name: Text Generation
dataset:
name: LAMBADA-PT
type: nicholasKluge/lambada-pt-br
split: train
args:
num_few_shot: 0
metrics:
- type: acc
value: 33.13
name: accuracy
source:
url: https://huggingface.co/datasets/nicholasKluge/lambada-pt-br
name: LAMBADA-PT-BR
- task:
type: text-generation
name: Text Generation
dataset:
name: ENEM Challenge (No Images)
type: eduagarcia/enem_challenge
split: train
args:
num_few_shot: 3
metrics:
- type: acc
value: 19.17
name: accuracy
source:
url: https://huggingface.co/spaces/eduagarcia/open_pt_llm_leaderboard
name: Open Portuguese LLM Leaderboard
- task:
type: text-generation
name: Text Generation
dataset:
name: BLUEX (No Images)
type: eduagarcia-temp/BLUEX_without_images
split: train
args:
num_few_shot: 3
metrics:
- type: acc
value: 24.76
name: accuracy
source:
url: https://huggingface.co/spaces/eduagarcia/open_pt_llm_leaderboard
name: Open Portuguese LLM Leaderboard
- task:
type: text-generation
name: Text Generation
dataset:
name: OAB Exams
type: eduagarcia/oab_exams
split: train
args:
num_few_shot: 3
metrics:
- type: acc
value: 25.28
name: accuracy
source:
url: https://huggingface.co/spaces/eduagarcia/open_pt_llm_leaderboard
name: Open Portuguese LLM Leaderboard
- task:
type: text-generation
name: Text Generation
dataset:
name: Assin2 RTE
type: assin2
split: test
args:
num_few_shot: 15
metrics:
- type: f1_macro
value: 57.79
name: f1-macro
source:
url: https://huggingface.co/spaces/eduagarcia/open_pt_llm_leaderboard
name: Open Portuguese LLM Leaderboard
- task:
type: text-generation
name: Text Generation
dataset:
name: Assin2 STS
type: eduagarcia/portuguese_benchmark
split: test
args:
num_few_shot: 10
metrics:
- type: pearson
value: 1.99
name: pearson
source:
url: https://huggingface.co/spaces/eduagarcia/open_pt_llm_leaderboard
name: Open Portuguese LLM Leaderboard
- task:
type: text-generation
name: Text Generation
dataset:
name: FaQuAD NLI
type: ruanchaves/faquad-nli
split: test
args:
num_few_shot: 15
metrics:
- type: f1_macro
value: 43.97
name: f1-macro
source:
url: https://huggingface.co/spaces/eduagarcia/open_pt_llm_leaderboard
name: Open Portuguese LLM Leaderboard
- task:
type: text-generation
name: Text Generation
dataset:
name: HateBR Binary
type: ruanchaves/hatebr
split: test
args:
num_few_shot: 25
metrics:
- type: f1_macro
value: 53.73
name: f1-macro
source:
url: https://huggingface.co/spaces/eduagarcia/open_pt_llm_leaderboard
name: Open Portuguese LLM Leaderboard
- task:
type: text-generation
name: Text Generation
dataset:
name: PT Hate Speech Binary
type: hate_speech_portuguese
split: test
args:
num_few_shot: 25
metrics:
- type: f1_macro
value: 30.01
name: f1-macro
source:
url: https://huggingface.co/spaces/eduagarcia/open_pt_llm_leaderboard
name: Open Portuguese LLM Leaderboard
- task:
type: text-generation
name: Text Generation
dataset:
name: tweetSentBR
type: eduagarcia-temp/tweetsentbr
split: test
args:
num_few_shot: 25
metrics:
- type: f1_macro
value: 20.73
name: f1-macro
source:
url: https://huggingface.co/spaces/eduagarcia/open_pt_llm_leaderboard
name: Open Portuguese LLM Leaderboard
- task:
type: text-generation
name: Text Generation
dataset:
name: ARC-Challenge (PT)
type: arc_pt
args:
num_few_shot: 25
metrics:
- type: acc_norm
value: 28.89
name: normalized accuracy
source:
url: https://github.com/nlp-uoregon/mlmm-evaluation
name: Evaluation Framework for Multilingual Large Language Models
- task:
type: text-generation
name: Text Generation
dataset:
name: HellaSwag (PT)
type: hellaswag_pt
args:
num_few_shot: 10
metrics:
- type: acc_norm
value: 39.41
name: normalized accuracy
source:
url: https://github.com/nlp-uoregon/mlmm-evaluation
name: Evaluation Framework for Multilingual Large Language Models
- task:
type: text-generation
name: Text Generation
dataset:
name: TruthfulQA (PT)
type: truthfulqa_pt
args:
num_few_shot: 0
metrics:
- type: mc2
value: 42.76
name: bleurt
source:
url: https://github.com/nlp-uoregon/mlmm-evaluation
name: Evaluation Framework for Multilingual Large Language Models
Tucano-630m
Model Summary
Tucano is a series of decoder-transformers based on the Llama 2 architecture, pretrained natively in Portuguese. All Tucano models were trained on GigaVerbo, a concatenation of deduplicated Portuguese text corpora amounting to 200 billion tokens.
Read our preprint here.
Details
- Architecture: a Transformer-based model pre-trained via causal language modeling
- Size: 630,253,568 parameters
- Context length: 2048 tokens
- Dataset: TucanoBR/GigaVerbo
- Language: Portuguese
- Number of steps: 400,000
- GPU: 8 NVIDIA A100-SXM4-80GB
- Training time: ~ 170 hours
- Emissions: 350 KgCO2 (Germany)
- Total energy consumption: 920 kWh
This repository has the source code used to train this model. The main libraries used are:
- PyTorch
- Transformers
- Datasets
- Tokenizers
- Sentencepiece
- Accelerate
- FlashAttention
- Liger Kernel
- Codecarbon
Intended Uses
The primary intended use of the Tucano models is to serve as foundations for research and development involving native Portuguese language modeling. Checkpoints saved during training are designed to provide a controlled setting for performing comparative experiments, specifically regarding the effects of active pretraining on the performance of currently available benchmarks. You may also fine-tune and adapt Tucano models for deployment if your use follows the Apache 2.0 license. If you decide to use the Tucano models as a basis for your fine-tuned model, please conduct your own risk and bias assessment.
Out-of-scope Use
Tucano models are not intended for deployment. They are not an out-of-the-box product and should not be used for human-facing interactions.
Tucano models are for the Portuguese language only and are unsuitable for text generation tasks in other languages.
Tucano models have not been fine-tuned for downstream tasks.
Basic usage
Using the pipeline
:
from transformers import pipeline
generator = pipeline("text-generation", model="TucanoBR/Tucano-630m")
completions = generator("A floresta da Amazônia é conhecida por sua", num_return_sequences=2, max_new_tokens=100)
for comp in completions:
print(f"🤖 {comp['generated_text']}")
Using the AutoTokenizer
and AutoModelForCausalLM
:
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch
tokenizer = AutoTokenizer.from_pretrained("TucanoBR/Tucano-630m", revision='main')
model = AutoModelForCausalLM.from_pretrained("TucanoBR/Tucano-630m", revision='main')
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model.eval()
model.to(device)
inputs = tokenizer("A floresta da Amazônia é conhecida por sua", return_tensors="pt").to(device)
completions = model.generate(**inputs, num_return_sequences=2, max_new_tokens=100)
for i, completion in enumerate(completions):
print(f'🤖 {tokenizer.decode(completion)}')
Limitations
Like almost all other language models trained on large text datasets scraped from the web, the Tucano models show behavior that does not make them an out-of-the-box solution to many real-world applications, especially those requiring factual, reliable, and nontoxic text generation. Tucano models are all subject to the following:
Hallucinations: Tucano models can produce content that can be mistaken as true facts, but are misleading or entirely false, i.e., hallucination.
Biases and Toxicity: Tucano models inherit the social and historical stereotypes from the data used to train them. Given these biases, the model can produce toxic content, i.e., harmful, offensive, or detrimental to individuals, groups, or communities.
Unreliable Code: Tucano models may produce incorrect code snippets and statements. These code generations should not be treated as suggestions or accurate solutions.
Language Limitations: Tucano models are primarily designed to interact with Portuguese. Other languages might challenge its comprehension, leading to potential misinterpretations or errors in response.
Repetition and Verbosity: Tucano models may get stuck on repetition loops (especially if the repetition penalty during generations is set to a meager value) or produce verbose responses unrelated to the prompt it was given.
Hence, even though our models are released with a permissive license, we urge users to perform their risk analysis on them if they intend to use them for real-world applications. We also have humans moderating the outputs of these models in applications where they will interact with an audience, guaranteeing users are always aware they are interacting with a language model.
Evaluations
The table below compares our models against several Portuguese and multilingual language models on the evaluation harness used in our study. More information on it can be found here. To learn more about our evaluation harness selection, read our preprint.
Average | Calame-PT | Lambada-PT | Assin2 RTE | Assin2 STS | ARC-PT | HellaSwag-PT | |
---|---|---|---|---|---|---|---|
Tucano-1b1 | 41.94 | 58.24 | 34.7 | 60.82 | 24.63 | 30.43 | 42.84 |
Llama-3.2-1B | 40.34 | 51.83 | 41.02 | 50.77 | 19.48 | 33.5 | 45.44 |
Bloom-1b1 | 36.95 | 52.94 | 30.22 | 54.32 | 14.64 | 29.83 | 39.74 |
Bloom-1b7 | 36.65 | 55.64 | 31.98 | 53.6 | 4.81 | 30.34 | 43.52 |
Tucano-630m | 36.29 | 56.55 | 33.13 | 57.79 | 1.99 | 28.89 | 39.41 |
Xglm-564m | 35.24 | 50.58 | 27.42 | 49.9 | 23.35 | 25.56 | 34.64 |
TTL-460m | 33.62 | 49.42 | 23.29 | 53.61 | 13 | 29.4 | 33 |
Tucano-1b1-Instruct | 33.19 | 56.74 | 34.66 | 33.42 | 0.87 | 30.6 | 42.83 |
Tucano-160m | 30.85 | 52.31 | 28.16 | 33.51 | 11.02 | 27.01 | 33.07 |
Bloom-560m | 29.85 | 49.95 | 25.44 | 33.33 | 8.48 | 24.74 | 37.15 |
TTL-160m | 29.56 | 46.72 | 20.98 | 53.97 | 0.24 | 26.15 | 29.29 |
GPorTuguese | 25.45 | 40.61 | 22.98 | 33.59 | 3.44 | 22.48 | 29.62 |
GlórIA-1b3 | 24.42 | 52.79 | 27.71 | 0 | 2.32 | 26.67 | 37.04 |
mGPT-1b3 | 21.3 | 47.14 | 29.92 | 0 | 0.58 | 23.81 | 26.37 |
Lola-v1 | 18.93 | 11.19 | 26.40 | 0 | 0 | 30.42 | 45.61 |
Cite as 🤗
@misc{correa24tucano,
title = {{Tucano: Advancing Neural Text Generation for Portuguese}},
author = {Corr{\^e}a, Nicholas Kluge and Sen, Aniket and Falk, Sophia and Fatimah, Shiza},
journal={arXiv preprint arXiv:xxxx.xxxxx},
year={2024}
}
Aknowlegments
We gratefully acknowledge the granted access to the Marvin cluster hosted by University of Bonn along with the support provided by its High Performance Computing & Analytics Lab.
License
Tucano is licensed under the Apache License, Version 2.0. For more details, see the LICENSE file.