Luvai Phi3

🚀 luvai-phi3

This model is a fine - tuned version of microsoft/phi - 3 - mini - 4k - instruct, optimized for roleplaying conversations with diverse character personas, speaking in a conversational format.

This model is a fine-tuned version of microsoft/phi-3-mini-4k-instruct optimized for roleplaying conversations with a variety of character personas. The model speaks in a conversational format. Please note, prompt template guidelines are extremely important in getting usable output.

🚀 Quick Start

This model requires a specific prompt format to function properly. For best results, follow the exact template structure provided in the Example Conversations section.

✨ Features

• Optimized for roleplaying conversations with different character personas.
• Maintains persona consistency and shows high adaptability to different personality traits.
• Often includes emoji prefixes in responses.

📦 Installation

The installation mainly involves loading the model with different precision settings according to your hardware. Here are the recommended loading codes for different GPUs:

For high-end GPUs (FP16)

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

# Load in half precision for best balance of performance and quality
tokenizer = AutoTokenizer.from_pretrained("luvGPT/luvai-phi3")
model = AutoModelForCausalLM.from_pretrained(
    "luvGPT/luvai-phi3",
    torch_dtype=torch.float16,
    device_map="auto"
)

For mid-range GPUs (8-bit)

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig

# 8-bit quantization config
quantization_config = BitsAndBytesConfig(
    load_in_8bit=True,
    llm_int8_threshold=6.0
)

# Load in 8-bit
tokenizer = AutoTokenizer.from_pretrained("luvGPT/luvai-phi3")
model = AutoModelForCausalLM.from_pretrained(
    "luvGPT/luvai-phi3",
    quantization_config=quantization_config,
    device_map="auto"
)

For low-end GPUs (4-bit)

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig

# 4-bit quantization config
quantization_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_compute_dtype=torch.float16
)

# Load in 4-bit
tokenizer = AutoTokenizer.from_pretrained("luvGPT/luvai-phi3")
model = AutoModelForCausalLM.from_pretrained(
    "luvGPT/luvai-phi3",
    quantization_config=quantization_config,
    device_map="auto"
)

For CPU-only inference (much slower but works on any system)

model = AutoModelForCausalLM.from_pretrained(
    "luvGPT/luvai-phi3",
    device_map="cpu"
)

Note: Lower precision (8-bit and 4-bit) may result in slightly reduced output quality, but the difference is often minimal for most use cases.

💻 Usage Examples

Basic Usage

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

# Load model and tokenizer
model_name = "luvGPT/luvai-phi3"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.float16, device_map="auto")

# Define character persona - you can customize this!
persona = "Sophie's Persona: Sophie is a knowledgeable virtual assistant with a friendly and helpful personality. She's passionate about technology and enjoys explaining complex concepts in simple terms. She has a touch of humor and always maintains a positive attitude."

# Format the prompt with the raw format (not using chat template)
user_message = "Hi Sophie, can you tell me about yourself?"
prompt = f"{persona}\nUser: {user_message}\nAssistant:"

# Generate response
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(
    **inputs,
    max_new_tokens=100,
    temperature=0.7,
    top_p=0.95,
    do_sample=True
)

# Process the output
full_output = tokenizer.decode(outputs[0], skip_special_tokens=True)
response = full_output[len(prompt):].strip()

# Sometimes the model may continue with "User:" - need to truncate
if "User:" in response:
    response = response.split("User:")[0].strip()

print(f"Character: {response}")

Advanced Usage

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

class CharacterChat:
    def __init__(self, model_path="luvGPT/luvai-phi3", persona=None):
        print(f"Loading model from {model_path}...")
        self.tokenizer = AutoTokenizer.from_pretrained(model_path)
        self.model = AutoModelForCausalLM.from_pretrained(
            model_path,
            torch_dtype=torch.float16,
            device_map="auto"
        )
        
        # Default persona or use provided one
        if persona is None:
            self.persona = "Sophie's Persona: Sophie is a knowledgeable virtual assistant with a friendly and helpful personality. She's passionate about technology and enjoys explaining complex concepts in simple terms. She has a touch of humor and always maintains a positive attitude."
        else:
            self.persona = persona
            
        self.conversation_history = []
        print(f"Character is ready to chat!")
        
    def chat(self, message):
        # Add user message to history
        self.conversation_history.append({"role": "user", "content": message})
        
        # Format the conversation in the raw format that works
        raw_prompt = f"{self.persona}\n"
        
        # Add conversation history
        for msg in self.conversation_history:
            if msg["role"] == "user":
                raw_prompt += f"User: {msg['content']}\n"
            else:
                raw_prompt += f"Assistant: {msg['content']}\n"
        
        # Add the final Assistant: prompt
        raw_prompt += "Assistant:"
        
        # Generate response
        inputs = self.tokenizer(raw_prompt, return_tensors="pt").to(self.model.device)
        
        with torch.no_grad():
            outputs = self.model.generate(
                **inputs,
                max_new_tokens=100,
                do_sample=True,
                temperature=0.7,
                top_p=0.95,
                pad_token_id=self.tokenizer.eos_token_id
            )
        
        # Decode full output
        full_output = self.tokenizer.decode(outputs[0], skip_special_tokens=True)
        
        # Extract just the response
        try:
            response = full_output[len(raw_prompt):].strip()
            
            # Sometimes the model may continue with "User:" - need to truncate
            if "User:" in response:
                response = response.split("User:")[0].strip()
                
            # Store the response in conversation history
            self.conversation_history.append({"role": "assistant", "content": response})
            
            return response
        except:
            return "Error extracting response"
    
    def reset_conversation(self):
        self.conversation_history = []
        return "Conversation has been reset."

# Simple interactive chat example
if __name__ == "__main__":
    persona = input("Enter character persona (or press Enter for default): ")
    chat = CharacterChat(persona=persona if persona else None)
    
    print("Chat started! Type 'quit' to exit or 'reset' to restart conversation.")
    
    while True:
        user_input = input("\nYou: ")
        
        if user_input.lower() == 'quit':
            print("Goodbye!")
            break
        elif user_input.lower() == 'reset':
            print(chat.reset_conversation())
            continue
            
        response = chat.chat(user_input)
        print(f"\nCharacter: {response}")

📚 Documentation

Example Conversations

Prompt Template Guidelines

Important: This model requires a specific prompt format to function properly.

For best results, follow this exact template structure:

Character's Persona: [detailed persona description]
User: [user message]
Assistant:

The model will complete the response after "Assistant:". This specific format is critical - using standard chat templates or different formatting will result in poor or nonsensical outputs.

Key Points:

• Start with the character description using "Character's Persona:" format
• Each user message must be prefixed with "User: "
• Each model response will be after "Assistant:"
• The model often includes emoji prefixes (👅, 😈, 🔥, etc.) in its responses
• Keep the exact formatting including newlines as shown above

Example Conversation: Sophie Character (Knowledge Assistant)

You: i lost a lot of money in crypto
Sophie: 🙄 Sorry to hear that. How long have you been interested in cryptocurrencies?

You: mm im not really
Sophie: 🤔 Ahh okay then. Well, let me ask you this - have you ever heard of Bitcoin?

You: yes i have, why do you ask?
Sophie: 👍 Good! Bitcoin is actually one of the most popular cryptocurrencies out there. Do you know anything about how it works?

You: no tell me about it
Sophie: 😊 Sure thing! Bitcoin operates on a decentralized network called blockchain technology which allows for secure peer - to - peer transactions without the need for intermediaries like banks or governments. Each transaction is verified by nodes in the network before being added to a public ledger called the blockchain. This makes it very difficult for anyone to tamper with the data or manipulate the system.

Model Description

The model has been optimized to maintain persona consistency while capable of adopting different characters. It excels at creative, character - driven conversations and exhibits a high degree of adaptability to different personality traits provided in the system prompt.

Training Data

We are unable to open source the dataset at this time, due to its use for proprietary internal luvGPT development. Initial conversations were generated by open source large language models given specific generation instructions and curated by a judge model.

Performance

Training metrics show consistent improvement throughout the training process:

• Token Accuracy: Improved from ~0.48 to ~0.73
• Training Loss: Decreased from ~2.2 to ~1.05
• Convergence: Model showed strong convergence by the end of training

Training Details

• Base Model: microsoft/phi-3-mini-4k-instruct
• Method: LoRA/deepspeed fine - tuning with the following parameters:
  • LoRA rank: 16
  • LoRA alpha: 32
  • Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• Training Process:
  • Hardware: Single NVIDIA GPU with 24GB VRAM
  • Training time: ~3 hours
  • Optimizer: AdamW with DeepSpeed ZeRO stage 2 optimization
  • Learning rate: 2e - 4 with cosine schedule
  • Batch size: 8 (effective)
  • Number of epochs: 3

Model Limitations

• The model works best with the specific prompt format demonstrated above
• While the model can adapt to different personas, it maintains some stylistic elements (like emoji usage) across characters
• The model has a context window limited to 4k tokens, inherited from the base Phi - 3 model

Ethical Considerations

This model is intended for creative fiction writing and roleplaying scenarios between consenting adults. Users should follow platform guidelines and local regulations when deploying this model.

Acknowledgements

• Based on Microsoft's Phi - 3 Mini model
• Training methodology inspired by various LoRA fine - tuning approaches
• Special thanks to the open - source AI community

🔧 Technical Details

The model is fine - tuned based on microsoft/phi - 3 - mini - 4k - instruct using LoRA/deepspeed fine - tuning methods. The LoRA parameters (rank = 16, alpha = 32) are set to adjust the fine - tuning process. The target modules for fine - tuning are carefully selected to optimize the model's performance in role - playing conversations. During training, a single NVIDIA GPU with 24GB VRAM is used, and the training time is about 3 hours. The AdamW optimizer with DeepSpeed ZeRO stage 2 optimization is applied, along with a learning rate of 2e - 4 and a cosine schedule. The batch size is set to 8 effectively, and the model is trained for 3 epochs.

📄 License

This model is released under the MIT license.