Instructions to use 22h/open-cabrita3b with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use 22h/open-cabrita3b with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="22h/open-cabrita3b")

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("22h/open-cabrita3b")
model = AutoModelForCausalLM.from_pretrained("22h/open-cabrita3b", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use 22h/open-cabrita3b with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "22h/open-cabrita3b"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "22h/open-cabrita3b",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/22h/open-cabrita3b

SGLang

How to use 22h/open-cabrita3b with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "22h/open-cabrita3b" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "22h/open-cabrita3b",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "22h/open-cabrita3b" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "22h/open-cabrita3b",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use 22h/open-cabrita3b with Docker Model Runner:
```
docker model run hf.co/22h/open-cabrita3b
```
Browse Quantizations to use this model in llama.cpp, Ollama, LM Studio, or any compatible app.

The Cabrita model is a collection of continued pre-trained and tokenizer-adapted models for the Portuguese language. This artifact is the 3 billion size variant.

The weights were initially obtained from the open-llama project (https://github.com/openlm-research/open_llama) in the open_llama_3b option.

@misc{larcher2023cabrita,
      title={Cabrita: closing the gap for foreign languages}, 
      author={Celio Larcher and Marcos Piau and Paulo Finardi and Pedro Gengo and Piero Esposito and Vinicius Caridá},
      year={2023},
      eprint={2308.11878},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	35.54
AI2 Reasoning Challenge (25-Shot)	33.79
HellaSwag (10-Shot)	55.35
MMLU (5-Shot)	25.16
TruthfulQA (0-shot)	38.50
Winogrande (5-shot)	59.43
GSM8k (5-shot)	0.99

Downloads last month: 343

Model tree for 22h/open-cabrita3b

Adapters

2 models

Quantizations

2 models

Spaces using 22h/open-cabrita3b 4

Paper for 22h/open-cabrita3b

Cabrita: closing the gap for foreign languages

Paper • 2308.11878 • Published Aug 23, 2023 • 1

Evaluation results

normalized accuracy on AI2 Reasoning Challenge (25-Shot)
test set Open LLM Leaderboard

33.790
normalized accuracy on HellaSwag (10-Shot)
validation set Open LLM Leaderboard

55.350
accuracy on MMLU (5-Shot)
test set Open LLM Leaderboard

25.160
mc2 on TruthfulQA (0-shot)
validation set Open LLM Leaderboard

38.500
accuracy on Winogrande (5-shot)
validation set Open LLM Leaderboard

59.430
accuracy on GSM8k (5-shot)
test set Open LLM Leaderboard

0.990