""" """ import random import gradio import config from app_util import * user_simulator_doc = """\ The agent acts as user simulator. There are maily two types of user simulator: - prompt-based user-simulator (role-play) - model-based user-simulator This demo is a 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. assistant_simulator_doc = """\ The agent acts as assistant simulator. """ self_chat_doc = """\ Self-chat is a demo which make the model talk to itself. It is a combination of user simulator and response generator. """ 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(head=None) as demo: # Knowledge Distillation through Self Chatting # Distilling the Knowledge from LLM through Self Chatting # Generating Synthetic Data through Self Chat gr.HTML("""

Generating Synthetic Data via Self-Chat

""") 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"), likeable=True) # gr.Textbox("For faster inference, you can build locally with ") # ss 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("🔄 Regenerate", 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_doc) # 也叫 chat-assistant, with gradio.Tab("Response Generator"): with gr.Row(): input_text_2 = gr.Textbox(show_label=False, placeholder="Please type user 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(assistant_simulator_doc) # with gradio.Tab("User Simulator"): with gr.Row(): input_text_3 = gr.Textbox(show_label=False, placeholder="Please type assistant 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", scale=1, min_width=300): # 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_new_tokens = gr.Slider(minimum=1, maximum=4096, value=config.DEFAULT_MAX_NEW_TOKENS, step=1, label="Max New 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", ) # TODO: gr.State 不能通过API传参。 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: self-chat generate_btn.click(chat, [chatbot, history], outputs=[chatbot, history], show_progress="full") retry_btn.click(undo_generate, [chatbot, history], outputs=[chatbot, history], show_api=False) \ .then(chat, [chatbot, history], outputs=[chatbot, history], show_progress="full", show_api=False) undo_btn.click(undo_generate, [chatbot, history], outputs=[chatbot, history], show_api=False) clear_btn.click(reset_state, inputs=[system], outputs=[chatbot, history], show_api=False) ######## tab2: response-generator generate_btn_2.click(append_user_to_history, [input_text_2, chatbot, history], outputs=[chatbot, history], show_api=False) \ .then(generate_assistant_message, [chatbot, history], outputs=[chatbot, history], show_progress="full", show_api=False) retry_btn_2.click(undo_generate, [chatbot, history], outputs=[chatbot, history], show_api=False) \ .then(chat, [chatbot, history], outputs=[chatbot, history], show_progress="full", show_api=False) undo_btn_2.click(undo_generate, [chatbot, history], outputs=[chatbot, history], show_api=False) clear_btn_2.click(reset_state, inputs=[system], outputs=[chatbot, history], show_api=False) \ .then(reset_user_input, outputs=[input_text_2], show_api=False) ######## tab3: user-simulator generate_btn_3.click(append_assistant_to_history, [input_text_3, chatbot, history], outputs=[chatbot, history], show_api=False) \ .then(generate_user_message, [chatbot, history], outputs=[chatbot, history], show_progress="full", show_api=False) retry_btn_3.click(undo_generate, [chatbot, history], outputs=[chatbot, history], show_api=False) \ .then(chat, [chatbot, history], outputs=[chatbot, history], show_progress="full", show_api=False) undo_btn_3.click(undo_generate, [chatbot, history], outputs=[chatbot, history], show_api=False) clear_btn_3.click(reset_state, inputs=[system], outputs=[chatbot, history], show_api=False) \ .then(reset_user_input, outputs=[input_text_3], show_api=False) slider_max_new_tokens.change(set_max_new_tokens, inputs=[slider_max_new_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.load(lambda: gr.update(value=random.choice(system_list)), None, system, show_api=False) # demo.queue().launch(share=False, server_name="0.0.0.0", debug=True) # demo.queue().launch(concurrency_count=1, max_size=5) demo.queue().launch()