Murai 350M V0.1 Beta

🚀 Murai-350M-v0.1-beta

A text generation model optimized for parameter efficiency and performance.

🚀 Quick Start

Here is a simple guide on how to use the Murai-350M-v0.1-beta model for text generation.

from transformers import pipeline, AutoTokenizer, AutoModelForCausalLM

model_id = "DeepMount00/Murai-350M-v0.1-beta"

model = AutoModelForCausalLM.from_pretrained(model_id, trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained(model_id)

t_pipeline = pipeline(
    "text-generation",
    model=model,
    tokenizer=tokenizer,
    device_map="auto",
    return_full_text=True, 
    top_p = 0.95, 
    top_k = 50,
    repetition_penalty=1.2
)

SYSTEM_PROMPT = """Sei un assistente utile."""
TEMPERATURE = 0.1
MAX_NEW_TOKENS = 250

messages = [
    {"role": "system", "content": SYSTEM_PROMPT},
    {"role": "user", "content": """Scrivi una funzione python che somma due numeri"""},
]

conv_template = tokenizer.apply_chat_template(
        messages,
        tokenize=False,
        add_generation_prompt=True
    )

outputs = t_pipeline(
    conv_template,
    max_new_tokens=MAX_NEW_TOKENS,
    do_sample=True,
    temperature=TEMPERATURE,
    num_return_sequences=1,
)
print(outputs[0]["generated_text"])

💻 Usage Examples

Basic Usage

# The above code demonstrates the basic usage of the model for text generation.
# It initializes the model, tokenizer, and pipeline, and then generates text based on a given prompt.
from transformers import pipeline, AutoTokenizer, AutoModelForCausalLM

model_id = "DeepMount00/Murai-350M-v0.1-beta"

model = AutoModelForCausalLM.from_pretrained(model_id, trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained(model_id)

t_pipeline = pipeline(
    "text-generation",
    model=model,
    tokenizer=tokenizer,
    device_map="auto",
    return_full_text=True, 
    top_p = 0.95, 
    top_k = 50,
    repetition_penalty=1.2
)

SYSTEM_PROMPT = """Sei un assistente utile."""
TEMPERATURE = 0.1
MAX_NEW_TOKENS = 250

messages = [
    {"role": "system", "content": SYSTEM_PROMPT},
    {"role": "user", "content": """Scrivi una funzione python che somma due numeri"""},
]

conv_template = tokenizer.apply_chat_template(
        messages,
        tokenize=False,
        add_generation_prompt=True
    )

outputs = t_pipeline(
    conv_template,
    max_new_tokens=MAX_NEW_TOKENS,
    do_sample=True,
    temperature=TEMPERATURE,
    num_return_sequences=1,
)
print(outputs[0]["generated_text"])

🔧 Technical Details

This model uses a deep architecture optimized for parameter efficiency:

• Pre-norm architecture with RMSNorm
• Grouped Query Attention for memory efficiency
• SwiGLU activation for improved performance
• RoPE position encoding for better length generalization

📚 Documentation

Citation

@misc{deepmount_llm_2024,
  title={Deep LLM: A 350M Parameter Language Model with 42 Layers},
  author={MicheleMontebovi},
  year={2025},
  url={https://huggingface.co/DeepMount00/Murai-350M-v0.1-beta}
}

📄 License

This project is licensed under the Apache 2.0 license.