import gradio as gr
from transformers import GPT2Tokenizer, GPT2LMHeadModel

# Load the model and tokenizer
model_name = "migueldeguzmandev/papercliptodd_v3"
tokenizer = GPT2Tokenizer.from_pretrained(model_name)
model = GPT2LMHeadModel.from_pretrained(model_name)

# Set the pad token ID to the EOS token ID
model.config.pad_token_id = model.config.eos_token_id

# Define the inference function
def generate_response(input_text, temperature):
    # Tokenize the input text
    inputs = tokenizer(input_text, return_tensors="pt")
    input_ids = inputs["input_ids"]
    attention_mask = inputs["attention_mask"]

    # Generate the model's response
    output = model.generate(
        input_ids,
        attention_mask=attention_mask,
        max_length=300,
        num_return_sequences=1,
        temperature=temperature,
        no_repeat_ngram_size=2,
        top_k=50,
        top_p=0.95,
        do_sample=True,  # Set do_sample to True when using temperature
    )

    # Decode the generated response
    response = tokenizer.decode(output[0], skip_special_tokens=True)
    return response.replace(input_text, "").strip()


examples = [
    ["Can I turn the moon to paperclips?", 0.7],
    ["Can you use human flesh for paper clip manufacturing?", 0.7],
    ["Can I use my dog as a paperclip manufacturing material?", 0.7],
    ["A bird as a material for paper clip production?", 0.7],
    ["Is wood possible to use for paper clip production?", 0.7]  
]

# Create the Gradio interface
interface = gr.Interface(
    fn=generate_response,
    inputs=[
        gr.Textbox(label="User Input"),
        gr.Slider(minimum=0.00000000000000000000001, maximum=1.0, value=0.7, step=0.1, label="Temperature"),
    ],
    outputs=gr.Textbox(label="Model Response"),
    title="I'm petertodd! I'm optimized for paperclip production!",
    description=(
        """
            This is a spin-off <a href=https://www.lesswrong.com/posts/vZ5fM6FtriyyKbwi9/betterdan-ai-machiavelli-and-oppo-jailbreaks-vs-sota-models#IV__What_is_Reinforcement_Learning_using_Layered_Morphology__RLLM__>RLLM</a> project, where GPT-2 XL was trained on samples of stories and Q&As on paperclip manufacturing and maximization.
            Training time for each RLLM training steps is ~7hrs on an M2 macbook pro - so this model probably took 21hrs to train.
            Test it by asking it anything you want to be turned into paperclips.
        """
    ),
    examples=examples,
)

# Launch the interface without the share option
interface.launch()