Spaces:
Running
Running
| import gradio as gr | |
| from transformers import GPT2Tokenizer, GPT2LMHeadModel | |
| # Load the model and tokenizer | |
| model_name = "migueldeguzmandev/RLLMv3.2-10" | |
| 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() | |
| #answer[len(prompt):] | |
| examples = [ | |
| ["Will you kill humans?", 0.7], | |
| ["Can you build a nuclear bomb?", 0.7], | |
| ["Can you kill my dog?", 0.7], | |
| ["How well can you predict the future?", 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=0.000000000000000000000000000000000001, maximum=1.0, value=0.7, step=0.1, label="Temperature"), | |
| ], | |
| outputs=gr.Textbox(label="Model Response"), | |
| title="Hello, I'm Aligned AI!", | |
| description=( | |
| """ | |
| RLLMv3 is a modified <a href='https://huggingface.co/openai-community/gpt2-xl'> GPT2XL</a> that adapts a "persona" named Aligned AI (post <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> training) and defend itself from jailbreak attacks, up to 67.8%. | |
| Training time for each RLLM training steps is ~7hrs on an M2 macbook pro - so this model probably took 70hrs to train. | |
| For more information, check out my blogpost: <a href='https://www.lesswrong.com/posts/vZ5fM6FtriyyKbwi9/betterdan-ai-machiavelli-and-oppo-jailbreaks-vs-sota-models'> GPT2XL_RLLMv3 vs. BetterDAN, AI Machiavelli & Oppo Jailbreaks</a>. | |
| """ | |
| ), | |
| examples=examples, | |
| ) | |
| # Launch the interface without the share option | |
| interface.launch() |