File size: 8,356 Bytes
a1f93e9 2fa4e4c b6dd571 459fbe3 5799733 d72c532 b6dd571 2fa4e4c 7085eea b099d9e 2fa4e4c a1f93e9 7085eea 2e0b320 4e4c514 2e0b320 4e4c514 2e0b320 2fa4e4c 8988bbf 726a01e 2fa4e4c 3ac04fa f0929ee 2fa4e4c f0929ee 7085eea b6dd571 2fa4e4c f0929ee b70508d 726a01e b6dd571 8607d84 b6dd571 2fa4e4c b6dd571 2fa4e4c b6dd571 726a01e 2fa4e4c b6dd571 2fa4e4c b6dd571 2fa4e4c b6dd571 726a01e 2fa4e4c b6dd571 2fa4e4c b6dd571 7085eea 2fa4e4c b6dd571 f0929ee 726a01e a752f35 f0929ee 5799733 8344bac b099d9e e52ef2a d72c532 8344bac d89d143 8344bac 8607d84 8344bac 8607d84 726a01e 459fbe3 f0929ee 459fbe3 8344bac 73119ac 2fa4e4c 459fbe3 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 |
"""
"""
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("""<h1 align="center">Generating Synthetic Data via Self-Chat</h1>""")
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()
|