Unified-Character-Model-DeepSeek

Unified Character Model

This is a fine-tuned DeepSeek-R1-Distill-Qwen-1.5B model that can roleplay as multiple characters based on the prompt.

Model Information

• Base Model: deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B
• Training Method: QLoRA (Quantized Low-Rank Adaptation)
• Fine-tuning Technique: Instruction fine-tuning with character dialogue

Usage

Python Code

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch

# Load base model with appropriate configuration
base_model = AutoModelForCausalLM.from_pretrained(
    "deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B",
    device_map="auto",
    torch_dtype=torch.float16,
    trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained("deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B")

# Load adapter weights
model = PeftModel.from_pretrained(base_model, "WompWomp1/Unified-Character-Model-DeepSeek")

# Function to generate responses for different characters
def generate_character_response(character_name, conversation, max_tokens=100):
    # Format the prompt
    prompt = f"""You are a roleplaying AI. Your role is to respond as {character_name}. When given a conversation, only respond with what {character_name} would say next.\n\nPrevious conversation:\n```\n{conversation}\n```\n\nIt's now {character_name}'s turn to speak. Respond in {character_name}'s voice and personality."""

    # Generate response
    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
    outputs = model.generate(
        input_ids=inputs.input_ids,
        max_new_tokens=max_tokens,
        temperature=0.7,
        do_sample=True,
        top_p=0.92,
        repetition_penalty=1.2,
        no_repeat_ngram_size=3
    )

    # Extract only the generated response
    full_response = tokenizer.decode(outputs[0], skip_special_tokens=True)
    prompt_decoded = tokenizer.decode(inputs.input_ids[0], skip_special_tokens=True)
    return full_response[len(prompt_decoded):]

# Example usage
conversation = "User: Hello! How are you doing today?"
response = generate_character_response("Ami", conversation)
print(f"Ami: {response}")

response = generate_character_response("Sensei", conversation)
print(f"Sensei: {response}")

Tips for Reducing Repetition

If you notice the model repeating phrases, try these settings:

1. Increase repetition_penalty (try values between 1.1 and 1.5)
2. Use no_repeat_ngram_size (values between 2-4 work well)
3. Reduce temperature slightly (0.6-0.7 range often works best)
4. Use both top_p (0.9-0.95) and top_k (40-50) together