wehub-resource-sync
317 lines
17 KiB
Python
317 lines
17 KiB
Python
|
||
""" This example shows how to build a unique 'Dueling Q&A ChatBot' in which a question-generating model 'chats'
|
||
with a question-answering model using a selected context passage from the user.
|
||
|
||
The user provides input of selecting the context passage, and then typing "Go" on the prompt bar, and the
|
||
'dueling' bots take it from there ...
|
||
|
||
Please note that two models will be downloaded and cached locally on first use - so please expect 1-2 minutes
|
||
for the first run, and then much faster loads on subsequent tries.
|
||
|
||
This example uses Streamlit for the UI. If you are new to using Steamlit, to run this example:
|
||
|
||
1. `pip3 install streamlit`
|
||
|
||
2. to run, go to the command line: streamlit run "path/to/dueling_chatbot.py"
|
||
|
||
"""
|
||
|
||
import streamlit as st
|
||
from llmware.models import ModelCatalog
|
||
from llmware.gguf_configs import GGUFConfigs
|
||
|
||
GGUFConfigs().set_config("max_output_tokens", 500)
|
||
|
||
if "question_history" not in st.session_state:
|
||
st.session_state["question_history"] = []
|
||
|
||
|
||
# test passage pulled from CNBC news story on Tuesday, May 28, 2024
|
||
test_passage = ("OpenAI said Tuesday it has established a new committee to make recommendations to the "
|
||
"company’s board about safety and security, weeks after dissolving a team focused on AI safety. "
|
||
"In a blog post, OpenAI said the new committee would be led by CEO Sam Altman as well as "
|
||
"Bret Taylor, the company’s board chair, and board member Nicole Seligman. The announcement "
|
||
"follows the high-profile exit this month of an OpenAI executive focused on safety, "
|
||
"Jan Leike. Leike resigned from OpenAI leveling criticisms that the company had "
|
||
"under-invested in AI safety work and that tensions with OpenAI’s leadership had "
|
||
"reached a breaking point.")
|
||
|
||
|
||
@st.cache_resource
|
||
def load_question_model(temperature=0.5):
|
||
""" Loads the Question Generating Model. """
|
||
question_model = ModelCatalog().load_model("slim-q-gen-tiny-tool",
|
||
temperature=temperature,
|
||
sample=True)
|
||
return question_model
|
||
|
||
|
||
@st.cache_resource
|
||
def load_answer_model():
|
||
""" Loads the Answering Model. """
|
||
answer_model = ModelCatalog().load_model("bling-phi-3-gguf",temperature=0.0, sample=False)
|
||
return answer_model
|
||
|
||
|
||
def get_new_question(q_model, question_type, test_passage):
|
||
|
||
new_q = ""
|
||
max_tries = 10
|
||
tries = 0
|
||
|
||
while not new_q or new_q in st.session_state["question_history"]:
|
||
|
||
response = q_model.function_call(test_passage, params=[question_type], get_logits=False)
|
||
if response:
|
||
if "llm_response" in response:
|
||
if "question" in response["llm_response"]:
|
||
new_q = response["llm_response"]["question"]
|
||
if isinstance(new_q, list) and len(new_q) > 0:
|
||
new_q = new_q[0]
|
||
if new_q not in st.session_state["question_history"]:
|
||
st.session_state["question_history"].append(new_q)
|
||
break
|
||
|
||
tries += 1
|
||
if tries >= max_tries:
|
||
break
|
||
|
||
return new_q
|
||
|
||
|
||
def get_new_answer(question, ans_model, test_passage):
|
||
|
||
response = ans_model.inference(question, add_context=test_passage)
|
||
answer = response["llm_response"]
|
||
|
||
return answer
|
||
|
||
|
||
def get_input_passage(sample_passage_name, custom_passage_text):
|
||
|
||
if sample_passage_name != "None":
|
||
|
||
if sample_passage_name == "OpenAI":
|
||
|
||
return ("OpenAI said Tuesday it has established a new committee to make recommendations to the "
|
||
"company’s board about safety and security, weeks after dissolving a team focused on AI safety. "
|
||
"In a blog post, OpenAI said the new committee would be led by CEO Sam Altman as well as "
|
||
"Bret Taylor, the company’s board chair, and board member Nicole Seligman. The announcement "
|
||
"follows the high-profile exit this month of an OpenAI executive focused on safety, "
|
||
"Jan Leike. Leike resigned from OpenAI leveling criticisms that the company had "
|
||
"under-invested in AI safety work and that tensions with OpenAI’s leadership had "
|
||
"reached a breaking point. The name of the new committee is the AI Safety Committee.")
|
||
|
||
elif sample_passage_name == "Apple":
|
||
|
||
return ("Apple shares popped 5% to a new record high of around $203 per share on Tuesday, a day "
|
||
"after the company announced its long-awaited push into artificial intelligence at its annual "
|
||
"developer conference on Monday. Apple introduced a range of new AI features during the event, "
|
||
"including an overhaul of its voice assistant Siri, integration with OpenAI’s ChatGPT, "
|
||
"a range of writing assistance tools and new customizable emojis. The company pitched the "
|
||
"features as AI for the average person, though users will likely need to upgrade their "
|
||
"iPhones to access the tools. With Tuesday’s share move, Apple bested its previous record "
|
||
"from Dec. 14. The company’s developer conference came as a welcome sign for investors who "
|
||
"have been watching to see how Apple will capitalize on the ongoing AI boom. Analysts from "
|
||
"Morgan Stanley said Apple’s AI features strongly position the company with “the most "
|
||
"differentiated consumer digital agent.” Additionally, the analysts believe that the "
|
||
"features will drive consumers to upgrade their iPhones, which should “accelerate "
|
||
"device replacement cycles.” They said Apple will still have to deliver when the AI "
|
||
"features are first available in the fall, but they think the “building blocks are in "
|
||
"place for a return to growth and more sustained outperformance.")
|
||
|
||
elif sample_passage_name == "Los Angeles Lakers":
|
||
|
||
return ("The Lakers have finished better than seventh in the Western Conference standings just once "
|
||
"in the past 12 seasons (when they won the title in 2019-20). Their franchise player, "
|
||
"LeBron James, turns 40 in December. They have no salary-cap space unless James were "
|
||
"to leave in free agency (he has until June 29 to make a decision on his $51.4 million "
|
||
"player option). They have limited trade assets. They play in a super-competitive conference "
|
||
"where the teams behind them are upwardly mobile and aggressive and most of the teams in "
|
||
"front of them are going to continue to be good -- or get even better -- in the "
|
||
"immediate future. And they have a massive and highly demanding fanbase and are "
|
||
"under a constant microscope by the national media because they drive massive "
|
||
"audience engagement across the world. Vogel won a title in 2020. Darvin Ham reached the "
|
||
"Western Conference finals in 2023. Neither lasted longer than three seasons. No "
|
||
"Lakers coach has since Phil Jackson has lasted more than three seasons. These are some of "
|
||
"the factors Hurley undoubtedly was weighing before making his choice over the weekend. "
|
||
"It's hard to even quantify what would be considered a successful season for the Lakers "
|
||
"in 2024-25 without knowing what changes are made to the roster. Avoiding the play-in tournament, "
|
||
"frankly, would be a reasonable if challenging goal.")
|
||
|
||
elif sample_passage_name == "Buy Home":
|
||
|
||
return ("The price for owning a home is rising rapidly and not just the mortgage payments. "
|
||
"US homeowners are now paying an average of $18,118 a year on property taxes, homeowners’ "
|
||
"insurance, maintenance, energy and various other expenses linked to owning a home, "
|
||
"according to a new Bankrate study. That’s nearly the cost to buy a used car and represents "
|
||
"a 26% increase from four years ago when it cost $14,428 annually to own and maintain a home. "
|
||
"All of these variable expenses are on top of the fixed cost of a mortgage, including "
|
||
"property taxes, homeowners insurance, energy costs, internet, cable bills and "
|
||
"home maintenance. The findings are another reminder of how much more expensive life "
|
||
"has become since Covid-19. Many Americans would like to buy a home but have been unable to "
|
||
"because home prices have spiked to record highs and mortgage rates remain elevated. "
|
||
"The housing market is historically unaffordable. But even the ones fortunate enough to "
|
||
"have bought a home over the past few years are grappling with sticker shock over the cost "
|
||
"of maintaining it. The per-month cost of owning and maintaining a home has gone from "
|
||
"$1,202 a month in 2020 to $1,510 now, Bankrate found.")
|
||
|
||
elif sample_passage_name == "Vacation":
|
||
|
||
return ("Temperatures are rising. Hotel prices are exploding. And travelers are already behaving badly. "
|
||
"Welcome to another summer in Europe. From the headlines, things already look chaotic. "
|
||
"Famous sites are raising their entry fees. Hotel rooms are like gold dust. And the "
|
||
"dollar has slipped against both the pound and the euro. Oh, and there’s the small matter "
|
||
"of crowds. “There’s been a substantial increase on last year’s demand,” says Tom "
|
||
"Jenkins, CEO of the European Tourism Organisation, speaking about US travelers to "
|
||
"Europe. “2023 saw higher numbers than 2019, and this year we’re comfortably seeing more – "
|
||
"record volumes of Americans coming to Europe.” Kayla Zeigler agrees. As the owner of "
|
||
"Destination Europe, she is sending “record numbers” of clients to the continent this year. "
|
||
"Graham Carter, director of Unforgettable Travel, a tour operator with a 90% US "
|
||
"client base, says that many guests are finding the idea of Europe prohibitively "
|
||
"expensive this year. People are wondering, is Europe worth it?” he says. “It’s "
|
||
"booking up in advance and prices are quite high. There’s been such a huge demand "
|
||
"for travel in the past three years, and lots of places are pushing up prices.” "
|
||
"Is summer in Europe already a washout? According to the experts, that all depends "
|
||
"on what kind of sacrifices you’re prepared to make. A weak dollar First things first: "
|
||
"travelers from the US are already at a disadvantage due to a weak dollar. Against "
|
||
"the euro, $1 was worth around 91 or 92 euro cents as of June 5, at mid-market rates. "
|
||
"Sure, that’s better than the December 2020-January 2021 five-year low when the "
|
||
"dollar was hovering around 82 cents. But it’s also down from a year earlier, "
|
||
"when a dollar was worth about 95 euro cents – and it’s way down from last "
|
||
"September’s five-year high when it peaked at 1.04 euros, according to currency "
|
||
"conversion specialists Wise. For those traveling to the UK it’s a similar state of "
|
||
"affairs. This time last year, $1 netted travelers 80 pence. As of Wednesday, it was "
|
||
"78p – a fall from the September peak of nearly 83p. The dollar is also down, year on year, "
|
||
"against 11 more European currencies. From Bosnia to Bulgaria, Denmark to Iceland, "
|
||
"Poland to Romania and Sweden to Switzerland, travelers changing dollars will be worse off. "
|
||
"While a few cents to the dollar doesn’t sound much on a single transaction, the small "
|
||
"drops can make a difference on credit card bills on the return home. A 500 euro hotel "
|
||
"room equates to $543 at Friday’s mid-market exchange rate, where it would have been "
|
||
"$480 in September.")
|
||
|
||
elif sample_passage_name == "Taylor Swift":
|
||
|
||
return ("Taylor Swift stopped her concert in Edinburgh, Scotland, on Friday to help a fan. Swift was "
|
||
"in the middle of singing her “Midnights” song “Would’ve Could’ve Should’ve” when she noticed "
|
||
"a fan who was in distress. In a video that went viral on social media, the singer-songwriter "
|
||
"is seen requesting assistance for the fan. “We need help right in front of me, please, "
|
||
"right in front of me,” Swift sang while playing her guitar and keeping her eyes locked on "
|
||
"the fan. “Just gonna keep playing until we notice where it is. Swift continued strumming "
|
||
"her guitar while motioning over to the person in need of help.")
|
||
|
||
else:
|
||
return custom_passage_text
|
||
else:
|
||
return custom_passage_text
|
||
|
||
|
||
def ask_and_answer_game(source_passage, q_model="slim-q-gen-tiny-tool", number_of_tries=10, question_type="question",
|
||
temperature=0.5):
|
||
|
||
""" Shows a simple two model game of using q-gen model to generate a question, and then a second model
|
||
to answer the question generated. """
|
||
|
||
# this is the model that will generate the 'question'
|
||
q_model = ModelCatalog().load_model(q_model, sample=True, temperature=temperature)
|
||
|
||
# this will be the model used to 'answer' the question
|
||
answer_model = ModelCatalog().load_model("bling-phi-3-gguf")
|
||
|
||
questions = []
|
||
|
||
print(f"\nGenerating a set of questions automatically from the source passage.\n")
|
||
|
||
for x in range(0,number_of_tries):
|
||
|
||
response = q_model.function_call(source_passage, params=[question_type], get_logits=False)
|
||
|
||
if response:
|
||
if "llm_response" in response:
|
||
if "question" in response["llm_response"]:
|
||
new_q = response["llm_response"]["question"]
|
||
|
||
# only keep new questions
|
||
if new_q and new_q not in questions:
|
||
questions.append(new_q)
|
||
|
||
print(f"inference - {x} - response: {response}")
|
||
|
||
print("\nAnswering the generated questions\n")
|
||
for i, question in enumerate(questions):
|
||
|
||
print(f"\nquestion: {i} - {question}")
|
||
if isinstance(question, list) and len(question) > 0:
|
||
response = answer_model.inference(question[0], add_context=test_passage)
|
||
print(f"response: ", response["llm_response"])
|
||
|
||
return True
|
||
|
||
|
||
def ask_and_answer_dueling_bots_ui_app (input_passage):
|
||
|
||
question_model = "slim-q-gen-phi-3-tool"
|
||
answer_model = "bling-stablelm-3b-tool"
|
||
|
||
st.title(f"Ask and Answer Dueling Bots")
|
||
st.write(f"Asking the questions: {question_model}")
|
||
st.write(f"Answering the questions: {answer_model}")
|
||
|
||
question_model = load_question_model()
|
||
answer_model = load_answer_model()
|
||
|
||
with st.sidebar:
|
||
|
||
st.write("Today's Subject")
|
||
sample_passage = st.selectbox("sample_passage", ["OpenAI", "Apple", "Los Angeles Lakers", "Buy Home",
|
||
"Taylor Swift", "Vacation", "None"], index=0)
|
||
# custom_passage = st.text_area(label="Subject",value=input_passage,height=100)
|
||
custom_passage = test_passage
|
||
|
||
input_passage = get_input_passage(sample_passage,custom_passage)
|
||
|
||
mode = st.selectbox("mode", ["question", "boolean", "multiple choice"],index=0)
|
||
|
||
number_of_tries = st.selectbox("tries", [5,10,20],index=0)
|
||
|
||
st.write(input_passage)
|
||
|
||
# initialize chat history
|
||
if "messages" not in st.session_state:
|
||
st.session_state.messages = []
|
||
|
||
# display chat messages from history on app rerun
|
||
for message in st.session_state.messages:
|
||
with st.chat_message(message["role"]):
|
||
st.markdown(message["content"])
|
||
|
||
# accept user input
|
||
prompt = st.chat_input("Say 'Go' and the Bots will Start")
|
||
if prompt:
|
||
|
||
for x in range(0, number_of_tries):
|
||
|
||
with st.chat_message("user"):
|
||
|
||
new_question = get_new_question(question_model,mode, input_passage)
|
||
st.markdown(new_question)
|
||
|
||
with st.chat_message("assistant"):
|
||
|
||
new_answer = get_new_answer(new_question,answer_model, input_passage)
|
||
st.markdown(new_answer)
|
||
|
||
st.session_state.messages.append({"role": "user", "content": new_question})
|
||
st.session_state.messages.append({"role": "assistant", "content": new_answer})
|
||
|
||
return 0
|
||
|
||
|
||
if __name__ == "__main__":
|
||
|
||
ask_and_answer_dueling_bots_ui_app(test_passage)
|
||
|
||
|
||
|
||
|