Spaces:
Runtime error
Runtime error
Commit
•
6d88f6f
1
Parent(s):
91190fb
Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,155 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Cloning our Github repository to Google Files
|
2 |
+
!git clone https://github.com/annikaijak/deeplearning_assignment_4
|
3 |
+
|
4 |
+
from langchain.document_loaders import PyPDFDirectoryLoader
|
5 |
+
from langchain.text_splitter import RecursiveCharacterTextSplitter
|
6 |
+
from langchain.embeddings import HuggingFaceEmbeddings
|
7 |
+
from langchain.vectorstores import Chroma
|
8 |
+
import torch
|
9 |
+
from langchain import HuggingFacePipeline
|
10 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer, GenerationConfig, pipeline
|
11 |
+
from langchain.chains import RetrievalQA
|
12 |
+
from langchain import PromptTemplate
|
13 |
+
from textwrap import fill
|
14 |
+
import gradio as gr
|
15 |
+
import time
|
16 |
+
|
17 |
+
|
18 |
+
# Loading the PDF files from Google Files
|
19 |
+
loader = PyPDFDirectoryLoader("/content/deeplearning_assignment_4/data/PDF_Documents")
|
20 |
+
docs = loader.load()
|
21 |
+
|
22 |
+
# Splitting the text in smaller chunks
|
23 |
+
text_splitter = RecursiveCharacterTextSplitter(chunk_size=1024, chunk_overlap=64)
|
24 |
+
texts = text_splitter.split_documents(docs)
|
25 |
+
|
26 |
+
# Creating embeddings
|
27 |
+
embeddings = HuggingFaceEmbeddings(
|
28 |
+
model_name="thenlper/gte-large",
|
29 |
+
model_kwargs={"device": "cuda"},
|
30 |
+
encode_kwargs={"normalize_embeddings": True},
|
31 |
+
)
|
32 |
+
|
33 |
+
query_result = embeddings.embed_query(texts[0].page_content)
|
34 |
+
|
35 |
+
# Saving the embeddings in the Chroma database
|
36 |
+
db = Chroma.from_documents(texts, embeddings, persist_directory="db")
|
37 |
+
results = db.similarity_search("Transformer models", k=2)
|
38 |
+
|
39 |
+
# Loading the transformer model
|
40 |
+
MODEL_NAME = "mistralai/Mistral-7B-Instruct-v0.1"
|
41 |
+
|
42 |
+
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME, use_fast=True)
|
43 |
+
|
44 |
+
model = AutoModelForCausalLM.from_pretrained(
|
45 |
+
MODEL_NAME, torch_dtype=torch.float16, trust_remote_code=True, device_map="auto"
|
46 |
+
)
|
47 |
+
|
48 |
+
# Create a configuration for text generation based on the specified model name
|
49 |
+
generation_config = GenerationConfig.from_pretrained(MODEL_NAME)
|
50 |
+
|
51 |
+
# Set the maximum number of new tokens in the generated text to 1024.
|
52 |
+
# This limits the length of the generated output to 1024 tokens.
|
53 |
+
generation_config.max_new_tokens = 1024
|
54 |
+
|
55 |
+
# Set the temperature for text generation. Lower values (e.g., 0.0001) make output more deterministic, following likely predictions.
|
56 |
+
# Higher values make the output more random.
|
57 |
+
generation_config.temperature = 0.0001
|
58 |
+
|
59 |
+
# Set the top-p sampling value. A value of 0.95 means focusing on the most likely words that make up 95% of the probability distribution.
|
60 |
+
generation_config.top_p = 0.95
|
61 |
+
|
62 |
+
# Enable text sampling. When set to True, the model randomly selects words based on their probabilities, introducing randomness.
|
63 |
+
generation_config.do_sample = True
|
64 |
+
|
65 |
+
# Set the repetition penalty. A value of 1.15 discourages the model from repeating the same words or phrases too frequently in the output.
|
66 |
+
generation_config.repetition_penalty = 1.15
|
67 |
+
|
68 |
+
|
69 |
+
# Create a text generation pipeline using the initialized model, tokenizer, and generation configuration
|
70 |
+
text_pipeline = pipeline(
|
71 |
+
"text-generation",
|
72 |
+
model=model,
|
73 |
+
tokenizer=tokenizer,
|
74 |
+
generation_config=generation_config,
|
75 |
+
)
|
76 |
+
|
77 |
+
# Create a LangChain pipeline that wraps the text generation pipeline and set a specific temperature for generation
|
78 |
+
llm = HuggingFacePipeline(pipeline=text_pipeline, model_kwargs={"temperature": 0})
|
79 |
+
|
80 |
+
template_3 = """
|
81 |
+
<s>[INST] <<SYS>>
|
82 |
+
Act as a student counselor at Aalborg University Business School and answer the question at the end.
|
83 |
+
The answer should be about the master programs found in the provided documents ONLY.
|
84 |
+
The answer should be MAXIMUM 40 words.
|
85 |
+
Use the examples in {context} to generate the answer, without directly mentioning any of it.
|
86 |
+
|
87 |
+
<</SYS>>
|
88 |
+
|
89 |
+
{context}
|
90 |
+
|
91 |
+
N-shot prompting:
|
92 |
+
N-1
|
93 |
+
Q: How do I find out what masters degree I want to study
|
94 |
+
A: To determine which master's degree you would like to study, you should consider which business-related modules are within your interest, which modeules from the bachelor's degree did you find intresting?
|
95 |
+
|
96 |
+
N-2
|
97 |
+
Q: I liked the modules [input] in the bachelor, what masters could be relevant for me?
|
98 |
+
A: Based on your interests in [input], it may be beneficial to consider studying [output].
|
99 |
+
The curriculum for this program includes several modules that align with your
|
100 |
+
interests.
|
101 |
+
|
102 |
+
ReAct prompting:
|
103 |
+
Q: "how do i find out what masters degree i want to study"
|
104 |
+
A: “To determine which master's degree you would like to study, you should consider which business-related modules are within your interest,
|
105 |
+
which modules from the bachelor's degree did you find interesting?
|
106 |
+
Q: "I liked macro economics and organisation"
|
107 |
+
A: “Based on your interests in macroeconomics and organizations, it may be
|
108 |
+
beneficial to consider studying the Master of Science (MSc) in Economics and
|
109 |
+
Business Administration (Finance) program at Aalborg University Business School.
|
110 |
+
The curriculum for this program includes several modules that align with your
|
111 |
+
interests, such as "Network Theory and Analysis" and "Data-Driven Business
|
112 |
+
Modeling and Strategy". These modules cover topics related to macroeconomics and
|
113 |
+
organizational behavior, providing you with valuable insights and skills that
|
114 |
+
could help you achieve your career goals. Additionally, the program offers an
|
115 |
+
application-focused approach, allowing you to apply your knowledge to real-world
|
116 |
+
problems and develop practical solutions.”
|
117 |
+
Feedback: The advice should focus on unique modules in the 1st and 2nd semester for each master, as the 3rd semester modules are elective options for all masters.
|
118 |
+
|
119 |
+
{question} [/INST]
|
120 |
+
"""
|
121 |
+
|
122 |
+
prompt_3 = PromptTemplate(template=template_3, input_variables=["context", "question"])
|
123 |
+
|
124 |
+
|
125 |
+
qa_chain_3 = RetrievalQA.from_chain_type(
|
126 |
+
llm=llm,
|
127 |
+
chain_type="stuff",
|
128 |
+
retriever=db.as_retriever(search_kwargs={"k": 2}),
|
129 |
+
return_source_documents=True,
|
130 |
+
chain_type_kwargs={"prompt": prompt_3},
|
131 |
+
)
|
132 |
+
|
133 |
+
def reply_bot(txt):
|
134 |
+
bot_result = qa_chain_3(txt)
|
135 |
+
return (bot_result["result"].strip())
|
136 |
+
|
137 |
+
bot_name = "Master Supervisor"
|
138 |
+
|
139 |
+
with gr.Blocks() as demo:
|
140 |
+
gr.Markdown("### Master's Degree Program Advisor")
|
141 |
+
gr.Markdown("I can help you find the master's degree program that's right for you. Ask me any question related to choosing a master's program.")
|
142 |
+
|
143 |
+
chatbot = gr.Chatbot()
|
144 |
+
msg = gr.Textbox()
|
145 |
+
clear = gr.ClearButton([msg, chatbot])
|
146 |
+
|
147 |
+
def reply_bot(message, chat_history):
|
148 |
+
bot_result = qa_chain_3(message)
|
149 |
+
chat_history.append((message, (bot_result["result"].strip()))),
|
150 |
+
time.sleep(2),
|
151 |
+
return "", chat_history
|
152 |
+
|
153 |
+
msg.submit(reply_bot, [msg, chatbot], [msg, chatbot])
|
154 |
+
|
155 |
+
demo.queue().launch(share=True)
|