Instructions to use BrCamp/bee-350m-pt-base-15b with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use BrCamp/bee-350m-pt-base-15b with Transformers:

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

pipe = pipeline("text-generation", model="BrCamp/bee-350m-pt-base-15b")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("BrCamp/bee-350m-pt-base-15b")
model = AutoModelForCausalLM.from_pretrained("BrCamp/bee-350m-pt-base-15b", device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use BrCamp/bee-350m-pt-base-15b with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "BrCamp/bee-350m-pt-base-15b"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "BrCamp/bee-350m-pt-base-15b",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/BrCamp/bee-350m-pt-base-15b

SGLang

How to use BrCamp/bee-350m-pt-base-15b 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 "BrCamp/bee-350m-pt-base-15b" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "BrCamp/bee-350m-pt-base-15b",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

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 "BrCamp/bee-350m-pt-base-15b" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "BrCamp/bee-350m-pt-base-15b",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use BrCamp/bee-350m-pt-base-15b with Docker Model Runner:
```
docker model run hf.co/BrCamp/bee-350m-pt-base-15b
```

Bee-350M PT — variante 15B (fork de decaimento)

Este e' o gemeo experimental de BrCamp/bee-350m-pt-base. Mesma arquitetura, mesmo corpus, mesmo tokenizador, mesma forma de schedule — a unica coisa que difere e' o total de tokens: 15,00B aqui contra 21,75B no principal.

Ele existe porque foi bifurcado do checkpoint do run principal no passo 165.000 (com estado do Adam e posicao no dado), decaindo dali ate 15,00B enquanto o principal seguia no plato.

Por que ele importa

modelo	tokens	tok/param	bpb
bee-350m-pt-base	21,75B	63	0,8207
este (15B)	15,00B	43	0,8223
bee-150m-pt-base	21,75B	143	0,8438

⭐ 45% mais dado renderam 0,19% de bpb. Ja ir de 151M para 345M parametros rendeu 2,76%. Nesta arquitetura o volume de tokens satura por volta de 43 tok/param — e a escala continua pagando.

⭐ E o mesmo modelo, nos mesmos 15,00B tokens, mede 0,9167 se colhido no plato do WSD e 0,8223 decaido: o decaimento de LR sozinho vale 10,3% de bpb. Comparar marcos intermediarios de modelos com schedules diferentes mede o schedule, nao o modelo.

Para que serve

Como modelo, ele e' ligeiramente pior que o principal — use o principal. Este aqui serve para reproduzir a ablacao: e' o unico ponto que permite isolar o efeito do volume de tokens com todo o resto fixo.

Detalhes: bpb medido.

Limitacoes

As mesmas do principal: modelo base, nao segue instrucoes, nao conversa, nao usa ferramentas; 345M parametros e 2048 de contexto alucinam fatos com facilidade; herda os vieses da web em portugues. bpb mede modelagem de linguagem, nao fluencia de resposta.

Downloads last month: 271

Safetensors

Model size

0.3B params

Tensor type

F32