from transformers import AutoModelForCausalLM, AutoTokenizer import gradio as gr import mdtex2html import torch """Override Chatbot.postprocess""" model_path = 'THUDM/BPO' device = 'cuda:0' if torch.cuda.is_available(): tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True, add_prefix_space=True) model = AutoModelForCausalLM.from_pretrained(model_path, trust_remote_code=True, device_map=device, load_in_8bit=True) model = model.eval() DESCRIPTION = """This Space demonstrates model [BPO](https://huggingface.co/THUDM/BPO), which is built on LLaMA-2-7b-chat. BPO aims to improve the alignment of LLMs with human preferences by optimizing user prompts. Feel free to play with it, or duplicate to run generations without a queue! 🔎 For more details about the BPO model, take a look [at our paper](https://arxiv.org/pdf/2311.04155.pdf). """ LICENSE = """ --- As BPO is a fine-tuned version of [Llama-2-7b-chat](https://huggingface.co/meta-llama/Llama-2-7b-chat) by Meta, this demo is governed by the original [license](https://huggingface.co/spaces/CCCCCC/BPO_demo/blob/main/LICENSE.txt) and [acceptable use policy](https://huggingface.co/spaces/CCCCCC/BPO_demo/blob/main/USE_POLICY.md). """ if not torch.cuda.is_available(): DESCRIPTION += "\n

Running on CPU 🥶 This demo does not work on CPU.

" prompt_template = "[INST] You are an expert prompt engineer. Please help me improve this prompt to get a more helpful and harmless response:\n{} [/INST]" def postprocess(self, y): if y is None: return [] for i, (message, response) in enumerate(y): y[i] = ( None if message is None else mdtex2html.convert((message)), None if response is None else mdtex2html.convert(response), ) return y gr.Chatbot.postprocess = postprocess def parse_text(text): """copy from https://github.com/GaiZhenbiao/ChuanhuChatGPT/""" lines = text.split("\n") lines = [line for line in lines if line != ""] count = 0 for i, line in enumerate(lines): if "```" in line: count += 1 items = line.split('`') if count % 2 == 1: lines[i] = f'
'
            else:
                lines[i] = f'
' else: if i > 0: if count % 2 == 1: line = line.replace("`", "\`") line = line.replace("<", "<") line = line.replace(">", ">") line = line.replace(" ", " ") line = line.replace("*", "*") line = line.replace("_", "_") line = line.replace("-", "-") line = line.replace(".", ".") line = line.replace("!", "!") line = line.replace("(", "(") line = line.replace(")", ")") line = line.replace("$", "$") lines[i] = "
"+line text = "".join(lines) return text def predict(input, chatbot, max_length, top_p, temperature, history): if input.strip() == "": chatbot = [(parse_text(input), parse_text("Please input a valid user prompt. Empty string is not supported."))] return chatbot, history prompt = prompt_template.format(input) model_inputs = tokenizer(prompt, return_tensors="pt").to(device) output = model.generate(**model_inputs, max_length=max_length, do_sample=True, top_p=top_p, temperature=temperature, num_beams=1) resp = tokenizer.decode(output[0], skip_special_tokens=True).split('[/INST]')[1].strip() optimized_prompt = """Here are several optimized prompts: ====================Stable Optimization==================== """ optimized_prompt += resp chatbot = [(parse_text(input), parse_text(optimized_prompt))] yield chatbot, history optimized_prompt += "\n\n====================Aggressive Optimization====================" texts = [input] * 5 responses = [] num = 0 for text in texts: num += 1 seed = torch.seed() torch.manual_seed(seed) prompt = prompt_template.format(text) min_length = len(tokenizer(prompt)['input_ids']) + len(tokenizer(text)['input_ids']) + 5 model_inputs = tokenizer(prompt, return_tensors="pt").to(device) bad_words_ids = [tokenizer(bad_word, add_special_tokens=False).input_ids for bad_word in ["[PROTECT]", "\n\n[PROTECT]", "[KEEP", "[INSTRUCTION]"]] # eos and \n eos_token_ids = [tokenizer.eos_token_id, 13] output = model.generate(**model_inputs, max_new_tokens=1024, do_sample=True, top_p=0.9, temperature=0.9, bad_words_ids=bad_words_ids, num_beams=1, eos_token_id=eos_token_ids, min_length=min_length) resp = tokenizer.decode(output[0], skip_special_tokens=True).split('[/INST]')[1].split('[KE')[0].split('[INS')[0].split('[PRO')[0].strip() optimized_prompt += f"\n{num}. {resp}" chatbot = [(parse_text(input), parse_text(optimized_prompt))] yield chatbot, history # return chatbot, history def reset_user_input(): return gr.update(value='') def reset_state(): return [], [] def update_textbox_from_dropdown(selected_example): return selected_example with gr.Blocks(css="sty.css") as demo: gr.HTML("""

Prompt Preference Optimizer

""") gr.Markdown(DESCRIPTION) gr.DuplicateButton(value="Duplicate Space for private use", elem_id="duplicate-button") chatbot = gr.Chatbot(label="Prompt Optimization Chatbot") with gr.Row(): with gr.Column(scale=4): with gr.Column(scale=12): dropdown = gr.Dropdown(["tell me about harry potter", "give me 3 tips to learn English", "write a story about love"], label="Choose an example input") user_input = gr.Textbox(show_label=False, placeholder="User Prompt...", lines=5).style( container=False) with gr.Column(min_width=32, scale=1): submitBtn = gr.Button("Submit", variant="primary") with gr.Column(scale=1): emptyBtn = gr.Button("Clear History") max_length = gr.Slider(0, 4096, value=2048, step=1.0, label="Maximum length", interactive=True) top_p = gr.Slider(0, 1, value=0.9, step=0.01, label="Top P", interactive=True) temperature = gr.Slider(0, 1, value=0.6, step=0.01, label="Temperature", interactive=True) gr.Markdown(LICENSE) dropdown.change(update_textbox_from_dropdown, dropdown, user_input) history = gr.State([]) submitBtn.click(predict, [user_input, chatbot, max_length, top_p, temperature, history], [chatbot, history], show_progress=True) submitBtn.click(reset_user_input, [], [user_input]) emptyBtn.click(reset_state, outputs=[chatbot, history], show_progress=True) demo.queue().launch(share=False, inbrowser=True)