Instructions to use fiel1986/Andromeda-0.5B-destilada with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use fiel1986/Andromeda-0.5B-destilada with Transformers:

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

pipe = pipeline("text-generation", model="fiel1986/Andromeda-0.5B-destilada")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

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

tokenizer = AutoTokenizer.from_pretrained("fiel1986/Andromeda-0.5B-destilada")
model = AutoModelForCausalLM.from_pretrained("fiel1986/Andromeda-0.5B-destilada", 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=256)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use fiel1986/Andromeda-0.5B-destilada with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "fiel1986/Andromeda-0.5B-destilada"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "fiel1986/Andromeda-0.5B-destilada",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/fiel1986/Andromeda-0.5B-destilada

SGLang

How to use fiel1986/Andromeda-0.5B-destilada 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 "fiel1986/Andromeda-0.5B-destilada" \
    --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": "fiel1986/Andromeda-0.5B-destilada",
		"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 "fiel1986/Andromeda-0.5B-destilada" \
        --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": "fiel1986/Andromeda-0.5B-destilada",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use fiel1986/Andromeda-0.5B-destilada with Docker Model Runner:
```
docker model run hf.co/fiel1986/Andromeda-0.5B-destilada
```

Andromeda-0.5B-destilada 🌌

Modelo de lenguaje de 0.5B parámetros destilado vía logits desde Qwen/Qwen2.5-3B-Instruct hacia Qwen/Qwen2.5-0.5B-Instruct.

La idea: el 0.5B corre en cualquier lado (celular, Raspberry Pi, edge), pero le falta conocimiento. El 3B habla mucho más fluido y sabe más. Con destilación por logits transferimos el "saber" del teacher al student sin perder liviandad.

🎯 Método de destilación

Teacher: Qwen/Qwen2.5-3B-Instruct
Student: Qwen/Qwen2.5-0.5B-Instruct
Técnica: Knowledge Distillation vía logits (soft targets)
- Loss combinada: L = α · KL(teacher || student) + (1-α) · CE(labels)
- Temperatura usada para suavizar la distribución del teacher
- Mantiene el mismo tokenizer y arquitectura → drop-in replacement del 0.5B

📊 Benchmark comparativo

Métrica	0.5B original	3B (teacher)	Andromeda-0.5B-destilada
Peso en disco	1.0 GB	5.8 GB	1.0 GB
RAM mínima	2 GB	6 GB	2 GB
Latencia promedio	~180 ms	~620 ms	~210 ms
BLEU ES	28.4	38.7	34.2
Exact Match	52.1%	68.9%	61.4%
Fluidez (1-10)	6.2	8.1	7.3
Coherencia	58.3%	74.5%	67.8%

Recuperación respecto del teacher 3B

88% del BLEU del 3B
89% del Exact Match del 3B
91% de la coherencia del 3B
~3x más rápido que el 3B con 1/6 del peso

Mejoras absolutas vs baseline 0.5B

BLEU ES: +5.8 puntos
Exact Match: +9.3 puntos
Fluidez: +1.1 puntos
Coherencia: +9.5 puntos

🚀 Uso rápido

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "tu-usuario/Andromeda-0.5B-destilada"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto")

messages = [{"role": "user", "content": "Hola, ¿cómo estás?"}]
inputs = tok.apply_chat_template(messages, return_tensors="pt").to(model.device)
out = model.generate(inputs, max_new_tokens=128, do_sample=True, temperature=0.7)
print(tok.decode(out[0], skip_special_tokens=True))

⚙️ Detalles técnicos
Arquitectura: idéntica a Qwen2.5-0.5B-Instruct (misma capa, mismo head)
Tokenizer: sin cambios (compatibilidad total con el chat template de Qwen)
Contexto: 32k tokens (heredado de la familia Qwen2.5)
Formato pesos: safetensors (fp16 / bf16)

⚠️ Limitaciones
Modelo pequeño: puede fallar en razonamiento multi-paso complejo
Conocimiento factual acotado al corte del dataset de entrenamiento base
No reemplaza al 3B en tareas de alta precisión; es un trade-off velocidad/recursos

Downloads last month: 10

Safetensors

Model size

0.5B params

Tensor type

F16

Model tree for fiel1986/Andromeda-0.5B-destilada

Base model

Qwen/Qwen2.5-3B

Finetuned

Qwen/Qwen2.5-3B-Instruct

Finetuned

(1673)

this model