đ Cabra Llama-3 8B
The Cabra Llama-3 8B is an enhanced version of the Meta-Llama-3-8B-Instruct, refined using the Cabra 30k dataset. This model is specially optimized to understand and respond in Portuguese.
Try the model on our chat.
Explore our other models and datasets, and the Cabra Llama 3 70b.

đ Detailed Documentation
Base Model Details
Model: Meta-Llama-3-8B-Instruct
Meta developed and released the Llama 3 family of models, a collection of pre - trained and instruction - tuned text - generation models in sizes of 8B and 70B. The instruction - tuned Llama 3 models are optimized for dialogue use cases and outperform many open - source chat models in common industry benchmarks. Moreover, great care was taken to optimize utility and safety during their development.
Model Architecture: Llama 3 is an auto - regressive language model that uses an optimized transformer architecture. The tuned versions use supervised fine - tuning (SFT) and reinforcement learning with human feedback (RLHF) to align with human preferences for utility and safety.
Dataset: Cabra 30k
It is an internal dataset for fine - tuning. We will release it soon.
Quantization / GGUF
We have placed various quantized versions (GGUF) in the "quantanization" branch.
Example
<s> [INST] Tell me some interesting facts about BrasĂlia. [/INST]
BrasĂlia is the capital of Brazil and was designed by Brazilian architect Oscar Niemeyer and urban planner LĂșcio Costa. Here are some interesting facts about BrasĂlia:
1. BrasĂlia was built in the 1960s. The city was constructed to replace the old capital of Brazil, Rio de Janeiro, which was becoming overcrowded.
2. The architecture of BrasĂlia is unique and is characterized by modernist and abstract forms. The city is divided into two main plans: the pilot plan, which is the central area of the city, and the ground plan, which is the residential area.
3. BrasĂlia is known for its iconic landmarks, such as the Metropolitan Cathedral of Our Lady of Aparecida, the PalĂĄcio da Alvorada, and the PalĂĄcio do Planalto, which are the workplaces of the President of Brazil.
4. The city is home to the National Stadium ManĂ© Garrincha, which hosts football games and has a capacity of 72,788 spectators. The stadium is named after the famous football player Garrincha, who was born in BrasĂlia.
5. BrasĂlia is a relatively new city and has a population of about 3 million people. Despite being relatively new, the city is known for its vibrant culture and is a center for tourism, business, and politics in Brazil.
</s>
đ» Usage
Currently, the model is intended for research purposes. Possible research areas and tasks include:
- Research on generative models.
- Investigation and understanding of the limitations and biases of generative models.
Commercial use is prohibited. Only for research. Contact us for more information.
Evaluations
Tasks |
Metric |
LLAMA3 Base Instruct Value |
Stderr |
Finetune Cabra Value |
Stderr |
assin2_rte |
f1_macro |
0.9091 |
0.0041 |
0.9036 |
0.0042 |
|
acc |
0.9093 |
0.0041 |
0.9036 |
0.0042 |
assin2_sts |
pearson |
0.7517 |
0.0074 |
0.6989 |
0.0082 |
|
mse |
0.5985 |
N/A |
0.6958 |
N/A |
bluex |
acc |
0.5786 |
0.0106 |
0.5786 |
0.0106 |
|
... |
... |
... |
... |
... |
enem |
acc |
0.7012 |
0.0070 |
0.7439 |
0.0067 |
faquad_nli |
f1_macro |
0.7516 |
0.0132 |
0.6988 |
0.0139 |
|
acc |
0.7938 |
0.0112 |
0.7508 |
0.0120 |
hatebr_offensive_binary |
f1_macro |
0.8699 |
0.0064 |
0.8528 |
0.0067 |
|
acc |
0.8700 |
0.0064 |
0.8536 |
0.0067 |
oab_exams |
acc |
0.5062 |
0.0062 |
0.4911 |
0.0062 |
portuguese_hate_speech_binary |
f1_macro |
0.5982 |
0.0120 |
0.5954 |
0.0120 |
|
acc |
0.5993 |
0.0119 |
0.5993 |
0.0119 |
đ Open Portuguese LLM Leaderboard Evaluation Results
Detailed results can be found here and on the đ Open Portuguese LLM Leaderboard
Metric |
Value |
Average |
69.42 |
ENEM Challenge (No Images) |
74.67 |
BLUEX (No Images) |
56.88 |
OAB Exams |
49.29 |
Assin2 RTE |
90.44 |
Assin2 STS |
69.85 |
FaQuAD NLI |
70.38 |
HateBR Binary |
85.05 |
PT Hate Speech Binary |
60.10 |
tweetSentBR |
68.08 |
đ License
This model is released under the cc - by - nc - 2.0 license.