""" """ import gradio import config from app_util import * system_list = [ "You are a helpful assistant.", "你是一个导游。", "你是一个英语老师。", "你是一个程序员。", "你是一个心理咨询师。", ] user_simulator_doc = """\ There are maily two types of user simulator: - prompt-based user-simulator (role-play) - model-based user-simulator In most cases, large language models (LLMs) are used to serve as assistant generator. Besides, it can also used as user simulator. """ survey = """\ ## knowledge distillation 知识蒸馏 Essentially, it is a form of model compression. ## distilling knowledge != knowledge distillation 知识的形式可以是 QA纯文本，也可以是 QA+概率。 ## 有不用概率的知识蒸馏吗？ """ with gr.Blocks() as demo: # Knowledge Distillation through Self Chatting # gr.HTML("""

Distilling the Knowledge from LLM through Self Chatting

""") with gr.Row(): with gr.Column(scale=5): system = gr.Dropdown( choices=system_list, value=system_list[0], allow_custom_value=True, interactive=True, label="System message", scale=5, ) chatbot = gr.Chatbot(show_copy_button=True, show_share_button=True, avatar_images=("assets/man.png", "assets/bot.png")) with gradio.Tab("Self Chat"): input_text_1 = gr.Textbox(show_label=False, placeholder="...", lines=10, visible=False) generate_btn = gr.Button("🤔️ Self-Chat", variant="primary") with gr.Row(): retry_btn = gr.Button("🔄 Retry", variant="secondary", size="sm", ) undo_btn = gr.Button("↩️ Undo", variant="secondary", size="sm", ) clear_btn = gr.Button("🗑️ Clear", variant="secondary", size="sm", ) # 🧹 Clear History (清除历史) # stop_btn = gr.Button("停止生成", variant="stop", visible=False) gr.Markdown( "Self-chat is a demo, which makes the model talk to itself. " "It is based on user simulator and response generator.", visible=True) with gradio.Tab("Response Generator"): with gr.Row(): input_text_2 = gr.Textbox(show_label=False, placeholder="Please type your input", scale=7) generate_btn_2 = gr.Button("Send", variant="primary") with gr.Row(): retry_btn_2 = gr.Button("🔄 Regenerate", variant="secondary", size="sm", ) undo_btn_2 = gr.Button("↩️ Undo", variant="secondary", size="sm", ) clear_btn_2 = gr.Button("🗑️ Clear", variant="secondary", size="sm", ) # 🧹 Clear History (清除历史) gr.Markdown("Response simulator is the most commonly used chatbot.") with gradio.Tab("User Simulator"): with gr.Row(): input_text_3 = gr.Textbox(show_label=False, placeholder="Please type your response", scale=7) generate_btn_3 = gr.Button("Send", variant="primary") with gr.Row(): retry_btn_3 = gr.Button("🔄 Regenerate", variant="secondary", size="sm", ) undo_btn_3 = gr.Button("↩️ Undo", variant="secondary", size="sm", ) clear_btn_3 = gr.Button("🗑️ Clear", variant="secondary", size="sm", ) # 🧹 Clear History (清除历史) gr.Markdown(user_simulator_doc) with gr.Column(variant="compact"): # with gr.Column(): model = gr.Dropdown( ["Qwen2-0.5B-Instruct", "llama3.1", "gemini"], value="Qwen2-0.5B-Instruct", label="Model", interactive=True, # visible=False ) with gr.Accordion(label="Parameters", open=True): slider_max_tokens = gr.Slider(minimum=1, maximum=config.MAX_SEQUENCE_LENGTH, value=config.DEFAULT_MAX_TOKENS, step=1, label="Max tokens") slider_temperature = gr.Slider(minimum=0.1, maximum=10.0, value=config.DEFAULT_TEMPERATURE, step=0.1, label="Temperature", info="Larger temperature increase the randomness") slider_top_p = gr.Slider( minimum=0.1, maximum=1.0, value=config.DEFAULT_TOP_P, step=0.05, label="Top-p (nucleus sampling)", ) slider_top_k = gr.Slider( minimum=1, maximum=200, value=config.DEFAULT_TOP_K, step=1, label="Top-k", ) history = gr.State([{"role": "system", "content": system_list[0]}]) # 有用信息只有个system，其他和chatbot内容重叠 system.change(reset_state, inputs=[system], outputs=[chatbot, history]) ######## tab1 generate_btn.click(generate, [chatbot, history], outputs=[chatbot, history], show_progress="full") retry_btn.click(undo_generate, [chatbot, history], outputs=[chatbot, history]) \ .then(generate, [chatbot, history], outputs=[chatbot, history], show_progress="full") undo_btn.click(undo_generate, [chatbot, history], outputs=[chatbot, history]) clear_btn.click(reset_state, inputs=[system], outputs=[chatbot, history]) ######## tab2 generate_btn_2.click(append_user, [input_text_2, chatbot, history], outputs=[chatbot, history]) \ .then(generate_assistant_message, [chatbot, history], outputs=[chatbot, history], show_progress="full") retry_btn_2.click(undo_generate, [chatbot, history], outputs=[chatbot, history]) \ .then(generate, [chatbot, history], outputs=[chatbot, history], show_progress="full") undo_btn_2.click(undo_generate, [chatbot, history], outputs=[chatbot, history]) clear_btn_2.click(reset_state, inputs=[system], outputs=[chatbot, history])\ .then(reset_user_input, outputs=[input_text_2]) ######## tab3 generate_btn_3.click(append_assistant, [input_text_3, chatbot, history], outputs=[chatbot, history]) \ .then(generate_assistant_message, [chatbot, history], outputs=[chatbot, history], show_progress="full") retry_btn_3.click(undo_generate, [chatbot, history], outputs=[chatbot, history]) \ .then(generate, [chatbot, history], outputs=[chatbot, history], show_progress="full") undo_btn_3.click(undo_generate, [chatbot, history], outputs=[chatbot, history]) clear_btn_3.click(reset_state, inputs=[system], outputs=[chatbot, history])\ .then(reset_user_input, outputs=[input_text_3]) slider_max_tokens.change(set_max_tokens, inputs=[slider_max_tokens]) slider_temperature.change(set_temperature, inputs=[slider_temperature]) slider_top_p.change(set_top_p, inputs=[slider_top_p]) slider_top_k.change(set_top_k, inputs=[slider_top_k]) # demo.queue().launch(share=False, server_name="0.0.0.0") # demo.queue().launch(concurrency_count=1, max_size=5) demo.queue().launch()