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

Libraries

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

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

pipe = pipeline("text-generation", model="BrCamp/bee-350m-pt-base")
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")
model = AutoModelForCausalLM.from_pretrained("BrCamp/bee-350m-pt-base", 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 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"
# 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",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

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

SGLang

How to use BrCamp/bee-350m-pt-base 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" \
    --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",
		"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" \
        --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",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

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

Bee-350M PT (base)

Modelo de linguagem de 345,4M parametros pre-treinado do zero em portugues, com tokenizador proprio de 32k. Nao e' fine-tune de nada: pesos aleatorios -> 21,75B tokens.

Este e' o modelo BASE. Ele nao segue instrucoes — para isso e' preciso SFT.

Resultado medido

bits-por-byte (bpb) num holdout limpo do fineweb-2 PT (400 documentos, sha256 2273c5e4…9663e028c). Menor e' melhor. bpb e' agnostico ao tokenizador, entao compara modelos com vocabularios diferentes de forma justa.

modelo	parametros	tokens	tok/param	bpb
Bee-350M (este)	345,4M	21,75B	63	0,8207
Bee-150M	151,2M	21,75B	143	0,8438

O Bee-350M supera o Bee-150M em 2,76% usando 3,3x menos tokens por parametro.

Como foi treinado


arquitetura	Llama-like, 32 camadas, d_model 960, 15q/5kv GQA, intermediate 2560
contexto	2048 tokens
corpus	21,75B tokens de portugues (fineweb-2 PT + dominio publico)
schedule	WSD — warmup 2%, plato em 55% do pico, decaimento `1-sqrt(t)` nos ultimos 20%
LR de pico	2,181e-03 (Step Law)
hardware	1x RTX 5090, 115,5 h
custo	~US$ 118

O que aprendemos treinando ele

Dois resultados medidos que valem mais que o modelo:

O decaimento de LR vale ~10% de bpb. O mesmo modelo, nos mesmos 15,00B tokens, mede 0,9167 no plato e 0,8223 decaido. Comparar marcos intermediarios de modelos com schedules diferentes mede o schedule, nao o modelo.
O volume de tokens saturou antes da escala. 15,00B -> 21,75B (+45% de dado) rendeu 0,19% de bpb. Ja 151M -> 345M parametros rendeu 2,76%.

Limitacoes

Modelo base: nao segue instrucoes, nao conversa, nao usa ferramentas.
345M parametros e 2048 de contexto — alucina fatos com facilidade.
Treinado predominantemente em web PT; herda os vieses dessa fonte.
bpb mede modelagem de linguagem. Fluencia de resposta e uso de ferramenta so' se medem pos-SFT, por execucao.

Uso

from transformers import AutoModelForCausalLM, AutoTokenizer

tok = AutoTokenizer.from_pretrained("BrCamp/bee-350m-pt-base")
modelo = AutoModelForCausalLM.from_pretrained("BrCamp/bee-350m-pt-base")
ids = tok("O Brasil e' um pais", return_tensors="pt")
print(tok.decode(modelo.generate(**ids, max_new_tokens=40)[0]))

Downloads last month: 1,240

Safetensors

Model size

0.3B params

Tensor type

F32