PDH Course E666 (2 PDH)

Introduction to ChatGPT

Helen Chen, Ph.D., PE, LEED AP

2023

PDH Online | PDH Center

5272 Meadow Estates Drive

Fairfax, VA 22030-6658
Phone: 703-988-0088
www.PDHonline.com
www.PDHcenter.org

An Approved Continuing Education Provider

www.PDHonline.com PDH Course E666 www.PDHcenter.org

Course Description

Chatbots, also known as “virtual assistants,” have become an increasingly common

way for businesses and organizations to interact with their customers and clients in
recent years. These chatbots are powered by artificial intelligence (AI) and natural
language processing (NLP) technology, allowing them to comprehend and respond
to user input in a conversational manner. In November 2022, OpenAI released
ChatGPT, a state-of-the-art language model, which was made available for users to
test and provide feedback on its capabilities and limitations.

In this course, you will gain a comprehensive understanding of ChatGPT and its
applications in natural language processing (NLP). You will learn about the basics of
NLP and transformer architecture, which is the basis for GPT. Additionally, you will
explore the limitations and ethical considerations associated with using GPT. By the
end of this course, you will have a good understanding of NLP and GPT, and be able
to apply this knowledge to your professional endeavors.

This course includes a multiple-choice quiz at the end, which is designed to enhance
your understanding of the course materials.

Learning Objectives

At the conclusion of this course, students will:

• Be able to use ChatGPT to generate texts and recommendations;

• Understand the basics of natural language processing (NLP) and transformer
architecture;
• Know the advantages and limitations of GPT;
• Understand the ethical considerations of using GPT.

©2023 Page 2 of 26
www.PDHonline.com PDH Course E666 www.PDHcenter.org

Table of Contents

Course Description ..................................................................................................... 2

Learning Objectives .................................................................................................... 2
Introduction .............................................................................................................. 4
AI-Related Terminology .............................................................................................. 4
Training .................................................................................................................... 6
What is ChatGPT? ....................................................................................................... 7
What Is a Prompt? ..................................................................................................... 8
How Does ChatGPT Work? ........................................................................................... 9
How Does ChatGPT Differ from Other Chatbots? ............................................................. 9
How to Use ChatGPT? ................................................................................................. 9
The Role of Prompts in ChatGPT Conversations ............................................................ 12
How to Write Effective Prompts? ................................................................................ 13
Natural Language Processing (NLP) ............................................................................ 15
Overview of GPT and Its Architecture .......................................................................... 16
Examples of How to Fine-tune GPT on Specific Tasks .................................................... 19
Advantages of GPT ................................................................................................... 21
Best Practices for Using GPT ...................................................................................... 22
Challenges and Limitations ........................................................................................ 23
Ethical Considerations and Limitations of Using GPT ..................................................... 24
Iterative Deployment of GPT and User Feedback .......................................................... 25
Conclusion and Future Outlook ................................................................................... 26

©2023 Page 3 of 26
www.PDHonline.com PDH Course E666 www.PDHcenter.org

Introduction to ChatGPT
Helen Chen, Ph.D., PE, LEED AP

Introduction
ChatGPT (short for "chat-based Generative
Pre-training Transformer") was launched as
a prototype on November 30, 2022, and
quickly garnered attention for its detailed
responses and articulate answers across
many domains of knowledge. This course
presents an overview of this powerful tool to
engineers and land surveyors, provides
guidance on how to use ChatGPT effectively,
and discusses the advantages and limitations of the generative pre-trained
transformer (GPT). As you work through the course materials, it is highly
recommended to visit the OpenAI website (https://chat.openai.com/chat) and
interact with ChatGPT to gain a firsthand understanding of its capabilities.

AI-Related Terminology
Understanding these terms will enhance your understanding of the course material:

Application Programming Interface (API) - A set of rules and protocols that

allows different software applications to communicate with each other. An API
defines how a developer can request data or services from an application, and how
that application can respond to those requests.

Artificial Intelligence (AI) - The branch of computer science that deals with
creating machines that can perform tasks that would normally require human
intelligence.

ChatGPT - A large language model chatbot developed by OpenAI based on GPT-

3.5. It has a remarkable ability to interact in conversational dialogue form and
provide human-like responses in a conversation.

Deep Learning (DL) - A subset of ML that uses deep neural networks to improve
the performance of AI systems.
©2023 Page 4 of 26
www.PDHonline.com PDH Course E666 www.PDHcenter.org

F1-Score - A measure of a model's accuracy that takes into account both precision
and recall. It is calculated as the harmonic mean of precision and recall, where the
best value is 1.0 and the worst value is 0.0.

Generative Models - A type of model that can generate new data, such as text or
images, based on training data.

Generative Pre-training Transformer (GPT) - A type of deep learning model

that is pre-trained on a large corpus of text data. It uses a transformer
architecture, which is a type of neural network that is particularly well-suited for
handling sequential data, such as text.

Hugging Face's Transformers - An open-source Python library that provides pre-

trained models and tokenizers for natural language processing tasks.

InstructGPT - A task-oriented language model developed by OpenAI. It is based

on the GPT-2 model and is fine-tuned to perform a specific task: generating
instructions.

Machine Learning (ML) - A type of AI that allows systems to learn and improve
from experience without being explicitly programmed.

Natural Language Processing (NLP) - A subfield of AI that deals with the

interaction between computers and human language.

Neural Networks - A type of machine learning model that is inspired by the

structure and function of the human brain.

Reinforcement Learning (RL) - A type of ML in which an agent learns to make

decisions by interacting with its environment and receiving rewards or penalties.

Supervised Learning - A type of machine learning where the model is trained on

a labeled dataset, where the correct output is provided for each input.

TensorFlow - An open-source software library for machine learning and deep

learning, developed by Google Brain Team.

Transformer - A deep learning model introduced in the paper "Attention Is All You
Need" by Google researchers in 2017. It is designed for natural language
processing tasks, such as language translation and text summarization.

Unsupervised Learning - A type of machine learning where the model is not

given any labeled data or any specific task to perform. Instead, the model is
presented with a dataset and must find patterns or relationships within the data on
its own.

©2023 Page 5 of 26
www.PDHonline.com PDH Course E666 www.PDHcenter.org

Training
OpenAI developed the ChatGPT model using a method known as Reinforcement
Learning from Human Feedback (RLHF), which is similar to the method used for
InstructGPT. However, there were slight variations in the data collection process.
Initially, OpenAI fine-tuned the model using supervised learning, where human
trainers engaged in conversations, assuming the roles of both the user and AI
assistant. The trainers were provided with suggestions generated by the model to
aid in composing their responses. The new dialogue dataset was then combined
with the InstructGPT dataset, which had been transformed into a dialogue format.

To create a reward model for reinforcement learning, OpenAI needed to collect

comparison data, which consisted of two or more model responses ranked by
quality. This data was collected by selecting conversations between the AI trainers
and the chatbot. Randomly chosen model-generated messages were used and
alternative completions were sampled, then trainers were asked to rank them. With
these reward models, OpenAI was able to fine-tune the model using Proximal Policy
Optimization. OpenAI performed several iterations of this process.

ChatGPT is fine-tuned from a model in the GPT-3.5 series, which finished training in
early 2022. ChatGPT and GPT 3.5 were trained on an Azure AI supercomputing
infrastructure.

Source: https://openai.com/blog/chatgpt/

©2023 Page 6 of 26
www.PDHonline.com PDH Course E666 www.PDHcenter.org

What is ChatGPT?
ChatGPT is a large language model developed by OpenAI that uses deep learning
techniques to generate text with a more natural flow, giving you the feeling of
conversing with a real person. When you type in a query, you should receive a
response that includes several sentences or paragraphs.

ChatGPT is based on the transformer architecture, which allows it to handle a wide

range of language tasks with minimal task-specific fine-tuning. It can be used for
tasks such as language translation, text summarization, and question answering,
and it has been used in various industries such as customer service, marketing, and
finance where understanding and generating human language is crucial.

Possible applications of ChatGPT in engineering include:

• Text generation for natural language interfaces in applications such as virtual

assistants, chatbots, and voice assistants.
• Text revisions or rewrites to improve the quality of technical writing.
• Language understanding for natural language processing tasks such as
sentiment analysis, text classification, and named entity recognition.
• Text summarization for extracting key information from large volumes of text.
• Language translation for multilingual applications.
• Text-to-speech synthesis for creating realistic and natural-sounding speech.
• Language modeling for improving the performance of other NLP tasks.

The following screen capture shows the home page of ChatGPT.

©2023 Page 7 of 26
www.PDHonline.com PDH Course E666 www.PDHcenter.org

What Is a Prompt?
A prompt is a phrase or individual keywords used as input to an AI tool like
ChatGPT. The tool then tries to analyze and understand the input and automatically
generates a response. So it’s important that you word your prompts in a way that
the tool can understand.

When writing prompts for ChatGPT there are some points to consider. First, the
prompt should be worded to include a clear task or question. Even though AI is
capable of processing incomplete or incorrect input, it is still helpful to give the tool
clear instructions.

Another important point is the choice of the right keywords. These should be as
precise and accurate to the question or task as possible. Only then can the AI tool
interpret the input correctly and deliver the desired results.

When you start working with ChatGPT, it is important that you write your prompts
correctly. A good prompt can help the tool work better and help you achieve your
goals. For this reason, we want to give you some examples of good and bad
prompts to give you a better idea of what to avoid.

Good prompts:

• "Generate a conversation between two friends discussing their plans for the
weekend."
• "Write a descriptive passage about a sunset on the beach."
• "Generate a script for a news segment about the latest breakthrough in
renewable energy."
• "Write a letter to a friend about the impact of the COVID-19 pandemic on
your life."
• "Write a short story about a robot who gains consciousness."

Bad prompts:

• "What is the capital of Australia?" (This is a question that can be easily

answered by a simple Google search and does not require a sophisticated
language model like GPT)
• "What's the weather like today?" (This is a task that is better suited for a
Google search rather than ChatGPT because ChatGPT is not connected to a
live weather API)
• "What is the meaning of life?" (This is a philosophical question that may not
have a clear answer and ChatGPT may not be able to provide a satisfactory
response)

©2023 Page 8 of 26
www.PDHonline.com PDH Course E666 www.PDHcenter.org

How Does ChatGPT Work?

ChatGPT works by using a transformer architecture, which is a type of neural
network that is particularly well-suited for processing sequential data such as text.
It uses self-attention mechanisms to weigh the importance of different words in a
sentence and generate a representation of the entire sentence. This allows the
model to understand the context and relationships between words in a sentence,
which is crucial for many NLP tasks such as language translation and question
answering.

The model is pre-trained on a large dataset of text, which allows it to learn the
statistical patterns and relationships between words and phrases in the language.
During the training process, the model is exposed to a large number of examples of
text, which allows it to learn how to generate and understand human language.
When given a prompt, the model uses this pre-trained knowledge to generate text
that is similar to the examples it was trained on.

The pre-training approach and the transformer architecture of GPT allow the model
to generate high-quality text, perform well on a wide range of language tasks with
minimal task-specific fine-tuning, and it's also able to generate human-like text.
However, it also requires large amounts of computational resources and may
generate biased or incorrect responses if the pre-training data contains biases.

How Does ChatGPT Differ from Other Chatbots?

Alexa by Amazon, Google Assistant, Siri by Apple, and Microsoft Cortana are some
of the other popular chatbots before OpenAI released ChatGPT in November 2022.
ChatGPT differs from these chatbots in a few ways. First, it uses a transformer-
based architecture, which allows it to handle long-term dependencies in language
more effectively. Additionally, it is pre-trained on a massive amount of text data,
which allows it to generate more human-like text and perform well on a wide range
of language tasks with minimal task-specific fine-tuning. Furthermore, it is more
powerful than other chatbots and can handle more complex and nuanced language
tasks.

How to Use ChatGPT?

ChatGPT tool is designed to give detailed responses to any inquiry you type - from
questions to statements. While the best results come from inputting a statement,
questions are also accepted. For example, if you type "explain a net-zero energy
building" you'll get a more detailed result than if you asked, "What is a net-zero

©2023 Page 9 of 26
www.PDHonline.com PDH Course E666 www.PDHcenter.org

energy building?" You can also get more specific and request a specific number of
paragraphs for an essay or a Wikipedia page.

©2023 Page 10 of 26
www.PDHonline.com PDH Course E666 www.PDHcenter.org

ChatGPT is not foolproof, though; if there is not enough data available, it may fill in
the gaps with incorrect information. OpenAI notes this is rare and that the tool also
currently has “limited knowledge of world events after 2021”, since it was trained
on data before the date.

ChatGPT can be used for many different tasks, for example:

• speech and text analysis

• translations
• explanations of complex issues
• writing stories and essays
• learn coding
• debugging code

OpenAI has released an official API for ChatGPT, which allows developers to easily
integrate the model into their applications and use it for a wide range of natural
language processing tasks, such as language translation, text summarization, and
question answering. The API provides access to the latest version of the model, and
enables fine-tuning on specific tasks and data sets. Additionally, the API allows for
easy integration with other services and platforms, such as cloud-based APIs and
web applications.

©2023 Page 11 of 26
www.PDHonline.com PDH Course E666 www.PDHcenter.org

The Role of Prompts in ChatGPT Conversations

At its most basic level, ChatGPT predicts text based on an input called a prompt.
Prompts play a crucial role in ChatGPT conversations as they provide the initial
input for the model to generate a response. The prompt acts as a seed for the
model, giving it context and direction for the conversation.

• Setting the context: The prompt sets the context for the conversation, giving
the model an understanding of the topic and any relevant background
information.
• Defining the scope: The prompt defines the scope of the conversation,
allowing the model to focus on specific aspects of the topic and provide
relevant information.
• Guiding the conversation: The prompt guides the conversation by providing a
starting point for the model's response, which can influence the direction of
the conversation.

©2023 Page 12 of 26
www.PDHonline.com PDH Course E666 www.PDHcenter.org

• Encouraging engagement: Clear and well-crafted prompts can encourage

engagement by making it easy for the user to understand the question and
respond with relevant information.

Crafting clear and concise prompts when using a language model like ChatGPT can
lead to a number of benefits:

• Increased accuracy: Clear and specific prompts make it easier for the model
to understand the intent behind the question and provide a more accurate
answer.
• Reduced response time: Concise prompts allow the model to quickly
understand the question and generate a response, reducing the overall time
it takes to get an answer.
• More efficient use of resources: When prompts are clear and specific, the
model is able to provide an answer with less computation, which can be more
efficient in terms of processing power and memory usage.
• Better understanding of context: Clear and specific prompts help the model
understand the context of the question, which can lead to more relevant and
useful answers.
• Better user experience: When prompts are clear and concise, it is easier for
the user to understand the question and the model's answer, which can lead
to a more positive and productive conversation.

In summary, crafting clear and concise prompts in ChatGPT conversation can help
to improve the overall accuracy and efficiency of the model, as well as the user's
experience.

How to Write Effective Prompts?

ChatGPT can tell bad jokes and write hilarious poems about your life, but it can also
help you do your job better. The catch: you need to help it do its job better, too, by
writing effective prompts. Here are some useful tips:

1. Offer context

Just like humans, AI does better with context. Think about exactly what you want
the AI to generate, and provide a prompt that's tailored specifically to that.

2. Include helpful information upfront

Let's say you want ChatGPT to write a speaker's introduction for yourself: how is
the AI supposed to know about you? It's not that smart (yet). But you can give it
the information it needs, so it can reference it directly. For example, you could copy

©2023 Page 13 of 26
www.PDHonline.com PDH Course E666 www.PDHcenter.org

your resume or LinkedIn profile and paste it at the top of your prompt. Likewise, if
you want the AI to summarize an article for you, you need to paste the full text of
the article above your prompt “Summarize the content from the above article with
5 bullet points.”

3. Give examples

Providing examples in the prompt can help the AI understand the type of response
you're looking for.

4. Tell it the length of the response you want

It's helpful to provide a word count for the response, so you don't get a 400-word
answer when you were looking for a sentence (or vice versa). You might even use a
range of acceptable lengths.

For example, if you want a 300-word response, you could provide a prompt like
"Write a 300-450-word summary of this article." This gives the AI the flexibility to
generate a response that's within the specified range. You can also use less precise
terms like "short" or "long."

5. Use some of these handy expressions

Sometimes it's just about finding the exact phrase that OpenAI will respond to.
Here are a few phrases that others have found work well with OpenAI to achieve
certain outcomes.

"Let's think step by step"

This makes the AI think logically and can be specifically helpful with math
problems.

"Thinking backwards"

This can help if the AI keeps arriving at inaccurate conclusions.

"In the style of [famous person]"

This will help match styles really well.

"As a [insert profession/role]"

This helps frame the bot's knowledge, so it knows what it knows—and what it
doesn't.

©2023 Page 14 of 26
www.PDHonline.com PDH Course E666 www.PDHcenter.org

Natural Language Processing (NLP)

Natural Language Processing (NLP) is a branch of Artificial Intelligence (AI) that
deals with the interaction between computers and humans using natural language.
The goal of NLP is to enable machines to understand, interpret, and generate
human language in a way that is both accurate and natural. NLP systems are used
to process and analyze large amounts of human language data, such as text,
speech, and video, in order to extract useful information and insights.

Here are some examples of common NLP tasks:

• Language Translation: A system that can translate text from one language to
another, for example, translating a news article from Spanish to English.
• Text Summarization: A system that can automatically generate a summary
of a long article or document, for example, summarizing a research paper
into a few key points.
• Sentiment Analysis: A system that can determine the sentiment or emotional
tone of a piece of text, for example, determining whether a customer review
of a product is positive or negative.
• Named Entity Recognition (NER): A system that can identify and extract
specific information such as people, places, and organizations from a text, for
example, identifying that "Barack Obama" is a person and "United States" is
a location in a news article.
• Part-of-Speech Tagging: A system that can identify and label the
grammatical parts of speech in a sentence, such as nouns, verbs, and
adjectives, for example, tagging "The cat sat on the mat" into "The/Det
cat/Noun sat/Verb on/prep the/Det mat/Noun"
• Parsing: A system that can analyze the syntactic structure of a sentence, for
example, understanding that "The cat sat on the mat" is a sentence with the
subject "cat" and object "mat"
• Question Answering (QA): A system that can understand the intent behind a
question and provide a relevant answer, for example, answering the question
"What is the capital of France?" with "Paris"
• Text Generation: A system that can create new text based on a given input,
for example, generating a continuation of a given sentence or generating a
new sentence based on a set of keywords.

These are just a few examples of the many tasks and applications that NLP systems
can be used for. It's important to note that NLP systems can be rule-based,
statistical, or neural network-based and the performance and accuracy may vary
depending on the task and the model used.

©2023 Page 15 of 26
www.PDHonline.com PDH Course E666 www.PDHcenter.org

NLP is a critical technology for many industries because it enables machines to

understand and generate human language in a natural and accurate way. Here are
a few examples of how NLP is used in different industries:

• Customer Service: NLP-powered chatbots and virtual assistants can provide

quick and accurate responses to customer inquiries, reducing wait times and
improving the overall customer experience. Additionally, sentiment analysis
can be used to understand customer feedback and improve products and
services.
• Marketing: NLP can be used to analyze large amounts of customer data to
gain insights into consumer behavior and preferences, which can inform
targeted marketing campaigns and product development. Additionally, text
summarization can be used to quickly analyze large amounts of social media
data and identify key trends.
• Finance: NLP can be used to analyze financial news articles and social media
posts to identify market sentiment and make predictions about stock prices.
Additionally, NLP can be used to extract relevant information from financial
documents such as earnings reports and SEC filings, which can inform
investment decisions.

Additionally, NLP is also being used in healthcare, e-commerce, and many other
industries. In healthcare, NLP-powered systems can extract useful information from
medical records and assist in diagnosis. In e-commerce, NLP-powered systems can
be used to extract information from product reviews and provide a summary.

Overview of GPT and Its Architecture

GPT, which stands for Generative Pre-training Transformer, is a large-scale neural
network-based language model developed by OpenAI. It uses unsupervised
learning to pre-train using a massive amount of text data and then fine-tune it on
specific tasks.

In traditional NLP models, the model is trained for a specific task using a labeled
dataset. For example, a sentiment analysis model is trained using a labeled dataset
of text samples with their corresponding sentiment labels (e.g. positive, negative).
GPT is different in the sense that it is pre-trained on a large corpus of text data
without any specific task in mind. This pre-training allows the model to learn
general language patterns and representations, which can then be fine-tuned for
specific tasks with a smaller labeled dataset.

GPT's architecture is based on the transformer architecture which is a neural

network architecture that was introduced in the 2017 paper "Attention Is All You
©2023 Page 16 of 26
www.PDHonline.com PDH Course E666 www.PDHcenter.org

Need" by Vaswani et al. The transformer architecture is particularly well-suited for

processing sequential data such as text. It uses self-attention mechanisms to weigh
the importance of different words in a sentence and generate a representation of
the entire sentence. This allows the model to understand the context and
relationships between words in a sentence, which is crucial for many NLP tasks such
as language translation and question answering.

The transformer architecture has two main components: the encoder and the
decoder. The encoder is responsible for encoding the input sequence into a fixed-
length representation, and the decoder is responsible for generating the output
sequence.

The encoder is composed of several layers of self-attention mechanisms and feed-

forward neural networks. The self-attention mechanism allows the model to weigh
the importance of different words in the input sequence and generate a fixed-length
representation of the entire sequence. The self-attention mechanism is calculated
by comparing each word in the input sequence with every other word, which allows
the model to understand the context and relationships between words in the input
sequence.

The decoder is also composed of several layers of self-attention mechanisms and

feed-forward neural networks. The decoder uses the fixed-length representation
generated by the encoder to generate the output sequence. The decoder also uses
an additional mechanism called masked self-attention, which only attends to the
input sequence before a certain position, in order to handle tasks like language
modeling where the model should not have access to the future tokens in the input.

In GPT, the transformer architecture is used to pre-train the model on a large

corpus of text data. The pre-trained model can then be fine-tuned on a specific task
using a smaller labeled dataset. The fine-tuning process can involve training only
the last layers of the model, or the entire model depending on the task and the size
of the labeled dataset.

The pre-training approach and the transformer architecture of GPT allow the model
to generate high-quality text, perform well on a wide range of language tasks with
minimal task-specific fine-tuning, and it's also able to generate human-like text.
However, it also requires large amounts of computational resources and may
generate biased or incorrect responses if the pre-training data contains biases.

GPT can be used in a variety of ways depending on the specific task and
requirements. Here are a few examples of the different ways that GPT can be used:

• Fine-tuning GPT on a specific task: One of the most common ways to use
GPT is to fine-tune it on a specific task, such as language translation, text

©2023 Page 17 of 26
www.PDHonline.com PDH Course E666 www.PDHcenter.org

summarization, or question answering, using a smaller labeled dataset. This

allows GPT to learn task-specific representations and improve its
performance on the specific task.
• Using GPT as a language model to generate text: GPT can also be used as a
language model to generate text. This can be used for text generation tasks
such as story generation, poetry generation, and automated content
creation. The user can provide a prompt and GPT generates text based on
the patterns it learned during pre-training.
• Using GPT as a feature extractor: GPT can also be used as a feature extractor
to extract representations of text that can be used for other tasks such as
sentiment analysis, text classification, and named entity recognition. These
representations can be obtained by encoding a text using GPT's encoder and
then using the encoded representation as input to another model that
performs the target task.
• GPT-based Applications: GPT can be used to build various applications such
as chatbots, virtual assistants, automated content creation, summarization
tools, Q&A systems and etc.
• GPT as a service: GPT can also be used as a service, for example, as a cloud-
based API, which allows developers to access the model's capabilities without
having to manage the computational resources required to run the model.

It's worth noting that the specific use-case, resources, and requirements should be
considered when choosing which way to use GPT, and it's also important to
evaluate the performance and limitations of the model on the target task.

©2023 Page 18 of 26
www.PDHonline.com PDH Course E666 www.PDHcenter.org

Examples of How to Fine-tune GPT on Specific Tasks

To fine-tune GPT on specific tasks, you can use the pre-trained GPT model and
adjust it to your desired task using a technique called transfer learning. This can be
done through the use of an API, such as the OpenAI API, which allows you to fine-
tune the model by providing it with a dataset for the specific task you want it to
perform. The API will then adjust the model's parameters based on the new data,
allowing it to perform the task more accurately.

There are several popular NLP libraries that provide pre-trained models and utilities
to fine-tune GPT on specific tasks. Here are a few examples of how to fine-tune GPT
on specific tasks using popular NLP libraries such as Hugging Face's Transformers
and TensorFlow:

• Hugging Face's Transformers: Hugging Face's Transformers library provides

pre-trained GPT models and utilities to fine-tune them on specific tasks. Here
is an example of fine-tuning GPT on a text classification task using the
Hugging Face's Transformers library:

from transformers import GPT2Tokenizer, GPT2ForSequenceClassification

from torch.utils.data import DataLoader
from torch.nn import CrossEntropyLoss

# Load the GPT2 tokenizer

tokenizer = GPT2Tokenizer.from_pretrained("gpt2")

# Load the GPT2 model

model = GPT2ForSequenceClassification.from_pretrained("gpt2")

# Prepare the dataset

data = ...
dataloader = DataLoader(data, batch_size=32, shuffle=True)

# Define the loss function

loss_fn = CrossEntropyLoss()

# Fine-tune the model

for epoch in range(num_epochs):
for input_ids, labels in dataloader:
# Forward pass
outputs = model(input_ids, labels=labels)
loss = outputs[0]
# Backward pass and optimization
loss.backward()
optimizer.step()
optimizer.zero_grad()

©2023 Page 19 of 26
www.PDHonline.com PDH Course E666 www.PDHcenter.org

• TensorFlow: TensorFlow also provides pre-trained GPT models and utilities to

fine-tune them on specific tasks. Here is an example of fine-tuning GPT on a
text classification task using TensorFlow:

import tensorflow as tf
from transformers import GPT2Tokenizer, GPT2ForSequenceClassification

# Load the GPT2 tokenizer

tokenizer = GPT2Tokenizer.from_pretrained("gpt2")

# Load the GPT2 model

model = GPT2ForSequenceClassification.from_pretrained("gpt2")

# Prepare the dataset

data = ...

# Define the loss function and optimizer

loss_fn = tf.keras.losses.SparseCategoricalCrossentropy()
optimizer = tf.keras.optimizers.Adam()

# Fine-tune the model

for epoch in range(num_epochs):
for input_ids, labels in data:
with tf.GradientTape() as tape:
# Forward pass
logits = model(input_ids, labels=labels)
loss_value = loss_fn(labels, logits)
# Backward pass and optimization
grads = tape.gradient(loss_value, model.trainable_weights)
optimizer.apply_grad

©2023 Page 20 of 26
www.PDHonline.com PDH Course E666 www.PDHcenter.org

Advantages of GPT
GPT has several advantages that make it a powerful and versatile language model.
Here are a few key advantages of GPT:

• Generating high-quality text: One of the main advantages of GPT is its ability
to generate high-quality and human-like text. This is due to its pre-training
on a large corpus of text data, which allows the model to learn general
language patterns and representations. This ability can be used for text
generation tasks such as story generation, poetry generation, and automated
content creation.
• Wide range of language tasks: GPT is trained on a wide range of text data,
which makes it capable of performing well on a wide range of language tasks.
It can be fine-tuned on specific tasks such as language translation, text
summarization, and question answering with minimal task-specific fine-
tuning. This is because the pre-training allows the model to learn general
language representations that can be useful for many tasks.
• Minimal task-specific fine-tuning: GPT is pre-trained on a large corpus of text
data, which means that it can be fine-tuned on specific tasks with a smaller
labeled dataset. This can save time and resources compared to traditional
models that require a large labeled dataset for each specific task.
• Smaller models: GPT models, especially GPT-3, have been trained on a large
amount of data and have billions of parameters, this allows them to have the
capability to perform well on a wide range of tasks with high quality and
accuracy.
• Flexible: GPT can be fine-tuned for a wide range of tasks, it also can be used
for different types of inputs such as text, speech, and video, and it can be
integrated into a wide range of applications such as chatbots, virtual
assistants, and automated content creation.

It's worth noting that GPT models can be quite large and computationally expensive
to fine-tune or use, so it's important to consider the specific use-case and available
resources when using GPT models.

©2023 Page 21 of 26
www.PDHonline.com PDH Course E666 www.PDHcenter.org

Best Practices for Using GPT

Following best practices when using GPT is essential for optimizing performance,
improving the user experience, and maintaining security and privacy. Some key
best practices for using GPT include:

• Evaluating the model's performance on the target task: Before deploying GPT
in production, it is important to evaluate the model's performance on the
target task to ensure that it meets the desired accuracy and quality
requirements. This can be done by fine-tuning the model on a dataset
specific to the target task and evaluating the model's performance using
metrics such as precision, recall, and F1-score.
• Managing resources: GPT models require a significant amount of
computational resources to run, it is important to manage the resources
properly, for example using cloud-based infrastructure to manage the
resources.
• Handling bias and errors: GPT models can produce biased or incorrect
responses if the data used to fine-tune the model is biased or if the model is
not fine-tuned on a diverse dataset. To mitigate this, it is important to use a
diverse and unbiased dataset to fine-tune the model, and to evaluate the
model's performance on a diverse set of inputs.
• Monitoring the model's performance: Once GPT is deployed in production, it
is important to monitor the model's performance to detect any issues or
errors that may occur. This can be done by logging the model's inputs and
outputs, and by monitoring the model's performance metrics.
• Improving the model's performance over time: Once GPT is deployed in
production, it is important to continuously improve the model's performance
by fine-tuning it on new data, and by incorporating feedback from users.

By following these best practices, you can ensure that GPT is used effectively and
efficiently in production, and that it meets the desired quality and accuracy
requirements.

©2023 Page 22 of 26
www.PDHonline.com PDH Course E666 www.PDHcenter.org

Challenges and Limitations

While GPT has many advantages, there are also several challenges and limitations
to consider when using this model. Here are a few key challenges of GPT:

• Computational resources: GPT models can be quite large and require

significant computational resources to fine-tune or use. This can be a
challenge for organizations with limited resources or for tasks that require
real-time processing.
• Biased or incorrect responses: GPT is pre-trained on a large corpus of text
data, which means that it may learn biases and inaccuracies present in the
data. This can result in biased or incorrect responses when the model is used
for specific tasks. This is particularly an issue when the data used for pre-
training is not diverse or has certain biases.
• Lack of control over the generated text: GPT generates text based on the
patterns it has learned during pre-training, and sometimes it can generate
text that is not coherent, not appropriate, or not aligned with the user's
expectations. This can be a problem for tasks that require a high degree of
control over the generated text, such as automated content creation.
• Lack of understanding of the context: GPT is good at understanding the
relationships between words in a sentence but it lacks the ability to
understand the context of the text. This can be a problem for tasks that
require understanding the meaning behind the text such as sentiment
analysis, sarcasm detection, and so on.
• Limited ability to handle structured data: GPT is mainly trained on
unstructured text data and it may not perform well on structured data such
as tables and graphs.

Besides the above challenges, GPT has the following limitations:

• The current GPT model (GPT-3) has a limitation of not having access to
information on recent world events that occurred after 2021. This is because
the model was trained on a dataset that was last updated in 2021, so it does
not have knowledge of events that occurred after that date.
• ChatGPT sometimes writes plausible-sounding but incorrect or nonsensical
answers. According to OpenAI, fixing this issue is challenging, as: (1) during
RL training, there’s currently no source of truth; (2) training the model to be
more cautious causes it to decline questions that it can answer correctly; and
(3) supervised training misleads the model because the ideal answer
depends on what the model knows, rather than what the human
demonstrator knows.

©2023 Page 23 of 26
www.PDHonline.com PDH Course E666 www.PDHcenter.org

• ChatGPT is sensitive to tweaks to the input phrasing or attempting the same

prompt multiple times. For example, given one phrasing of a question, the
model can claim to not know the answer, but given a slight rephrase, can
answer correctly.
• The model is often excessively verbose and overuses certain phrases, such
as restating that it’s a language model trained by OpenAI. These issues arise
from biases in the training data (trainers prefer longer answers that look
more comprehensive) and well-known over-optimization issues.
• Ideally, the model would ask clarifying questions when the user provided an
ambiguous query. Instead, OpenAI’s current models usually guess what the
user intended.
• While OpenAI has made efforts to make the model refuse inappropriate
requests, it will sometimes respond to harmful instructions or exhibit biased
behavior. OpenAI is using the Moderation API to warn or block certain types
of unsafe content, but they expect it to have some false negatives and
positives for now.

In order to overcome these challenges and limitations, it's important to carefully

evaluate the suitability of GPT for a specific task, to use diverse and unbiased data
for pre-training, to fine-tune the model with a smaller task-specific labeled dataset,
and to monitor the generated text for biases and inaccuracies.

Ethical Considerations and Limitations of Using GPT

There are several ethical considerations and limitations to using GPT, which include:

• Bias: GPT models are trained on large amounts of text data, which may
contain biases that are reflected in the model's outputs. This can result in
GPT producing biased or discriminatory responses, particularly when the data
used to fine-tune the model is not diverse and inclusive.
• Misinformation: GPT can generate high-quality text, but it can also generate
text that is factually incorrect or misleading. This is especially problematic
when GPT is used to generate news articles, product descriptions, or other
types of content that are meant to inform the public.
• Privacy: GPT models require large amounts of data to train, which may
include sensitive personal information. This raises privacy concerns and it is
important to ensure that the data used to train GPT models is properly
anonymized and that the models are not used to make decisions that impact
people's lives.
• Dependence: GPT models can generate high-quality text, but they can also
create a dependency on the technology, particularly when it comes to tasks
©2023 Page 24 of 26
www.PDHonline.com PDH Course E666 www.PDHcenter.org

such as content creation. This can lead to a loss of creativity, and a lack of
critical thinking.
• Decision making: GPT models can be used to make decisions, for example, to
generate responses for customer service chatbot or to generate investment
recommendations. However, GPT models are not perfect and they can make
mistakes, it is important to ensure that the decisions made by GPT models
are auditable and that there are mechanisms in place to detect and correct
errors.

It's important to be aware of these ethical considerations and limitations when

using GPT and to use the technology responsibly. Additionally, it is important to
evaluate the performance and limitations of the model on the target task, and to
make sure the data used for fine-tuning the model is diverse and unbiased. Also, it
is important to monitor and evaluate the performance of the model after it is
deployed in production and to continuously improve the model's performance by
fine-tuning it on new data, and by incorporating feedback from users.

Iterative Deployment of GPT and User Feedback

The current release of ChatGPT is the latest step in OpenAI’s iterative deployment
of increasingly safe and useful AI systems. Many lessons from the deployment of
earlier models like GPT-3 and Codex have informed the safety mitigations in place
for this release, including substantial reductions in harmful and untruthful outputs
achieved by the use of reinforcement learning from human feedback (RLHF).

OpenAI recognizes that there are still limitations with their current models. They
intend to make frequent updates to the model in order to address such limitations.
Additionally, by making ChatGPT easily accessible to users, OpenAI hopes to gather
valuable feedback on issues that they may not have previously been aware of.

Users are encouraged to provide feedback on problematic model outputs through

the UI, as well as on false positives/negatives from the external content filter which
is also part of the interface. OpenAI is particularly interested in feedback regarding
harmful outputs that could occur in real-world, non-adversarial conditions, as well
as feedback that helps them uncover and understand novel risks and possible
mitigations.

©2023 Page 25 of 26
www.PDHonline.com PDH Course E666 www.PDHcenter.org

Conclusion and Future Outlook

ChatGPT is a powerful conversational language model developed by OpenAI. It is
designed to be able to generate natural-sounding and coherent responses to a wide
variety of prompts. It can respond to prompts about various topics, including pop
culture, personal interests, and technical subjects. The model can be fine-tuned on
a specific domain or task, to generate responses that are more accurate and
relevant to the conversation. ChatGPT can be used in a variety of applications,
including chatbots, virtual assistants, and customer service bots. It can also be
used to generate dialogue for video games, chat fiction, and other interactive
media. Additionally, it can be used to create automated content such as news
articles, blog posts, and social media posts. Its dialogue format makes it possible to
answer followup questions, admit its mistakes, challenge incorrect premises, and
reject inappropriate requests.

NLP and GPT are valuable tools that are revolutionizing the way we interact with
and understand human language. NLP provides the means to process and analyze
natural language, which is crucial in a wide range of industries such as customer
service, marketing, and finance. GPT, in particular, is a powerful language model
that is capable of generating high-quality text and performing well on a wide range
of language tasks with minimal task-specific fine-tuning.

However, the use of GPT and NLP also brings ethical considerations and limitations
that need to be taken into account, such as bias, misinformation, privacy,
dependence, and decision-making. To ensure the responsible use of GPT and NLP, it
is important to evaluate the performance and limitations of the model on the target
task, and to make sure the data used for fine-tuning the model is diverse and
unbiased. Additionally, it is important to monitor and evaluate the performance of
the model after it is deployed in production and to continuously improve the
model's performance by fine-tuning it on new data, and by incorporating feedback
from users.

Looking into the future, there is ongoing research to scale GPT to handle larger
inputs and outputs, to improve the explainability of GPT, to incorporate multi-modal
inputs, to develop GPT models for low-resource languages, and to incorporate GPT
in other AI applications. Additionally, further research is needed to address the
ethical considerations and limitations of GPT and NLP. With these advancements,
we can expect to see even more powerful and sophisticated NLP and GPT models
that can handle a wide range of language tasks and improve our ability to
understand and generate human language.

©2023 Page 26 of 26