A Comprehensive

Guide to Generative AI
This document aims to answer general questions on Generative
AI, Foundation Models, and OpenAI integration in DataRobot
DATA SHEE T | A Comprehensive Guide to Generative AI

CONTENTS

General Questions About Generative AI 3

What are Generative AI, Foundation Models, and LLMs 3
What are the capabilities of Foundation Models 3
What are the different types of Foundation models 5
How Foundation Models are developed 5
How Foundation Models can be used in business 6
What are the strength of Foundation Models 6
What are the weaknesses of Foundation Models 7
What are the risks and ethical concerns associated with the use of Foundation Models 7

DataRobot LLM integrations 8

How does DataRobot integrate with Azure OpenAI LLMs? 8
Does DataRobot share data with Azure OpenAI and Microsoft? 8
What kind of data is shared with Microsoft Azure for our Azure OpenAI integration? 8
Does DataRobot collect data from users? 8
Will the Azure OpenAI integration be on by default for customers in preview mode,
or will that need to be turned on?  9
Will the Azure OpenAI integration be made available for Single-Tenant SaaS
or Self-Managed Installations? 9

DataRobot Guideline on the use of integrated LLM models  9

Where can I find DataRobot guidelines on acceptable use of LLM integrated capabilities? 9
What is an acceptable prompt for DataRobot Notebook Code Assist? 9
What is an acceptable prompt for the DataRobot Insight App? 9

Best Practices on the use of LLMs 9

What are the best practices for prompt engineering? 9

2
DATA SHEE T | A Comprehensive Guide to Generative AI

GENERAL QUESTIONS ABOUT GENERATIVE AI

What are Generative AI, Foundation Models, and LLMs

Generative AI: Generative AI refers to a type of artificial intelligence that can create new content.
It’s designed to learn patterns in data and generate new content that fits the same patterns.
The most common use cases of generative AI include creating synthetic images, music, voiceovers,
and text. A popular form of generative AI is the generative adversarial network (GAN), which uses
two neural networks, a generator and a discriminator, to create and refine its output.

Foundation Model: Foundation models are a type of artificial intelligence model that is pre-trained
on a broad range of Internet text. They can understand and generate human language to a remarkable
degree. Examples of text foundation models are GPT-4 and PaLM. These models are referred
to as “foundation” because they can be fine-tuned for a wide variety of specific tasks, much like how
a foundation supports a range of structures above it.

LLM (Large Language Model): This term generally refers to a class of machine learning models that
learn to understand, generate, or translate human language. These models learn from large amounts
of text data and are capable of tasks such as answering questions, translating languages, summarizing
text, and more. Examples of such models are the Transformer-based models like BERT (Bidirectional
Encoder Representations from Transformers), GPT (Generative Pretrained Transformer), and others.

What are the capabilities of Foundation Models

Foundation Models, such as GPT-4 and DALL-E are versatile and capable of a wide range of tasks related
to understanding and generating human language. These models are trained on large corpora of text,
image, and audio data and can be fine-tuned for more specific tasks. Here are some examples
of the LLM capabilities:

Text Generation: These models are capable of generating human-like text. This can be used for marketing,
sales, or even for more practical purposes like drafting emails or other forms of communication.

Question Answering: Given a question, Foundation Models can generate an appropriate response
based on their understanding of the text. This is commonly used in customer service chatbots
and virtual assistants.

Translation: These models can translate text from one language to another. They have been trained on text
data in multiple languages and can thus understand and generate text in those languages.

Text Summarization: Foundation Models can take a long piece of text and create a shorter summary
that retains the most important points.

Information Extraction: Foundation models can extract structured information from unstructured
text data. For example, they can be used to extract names, dates, and places from a news article.

Completion and Prediction: Foundation Models can predict what comes next in a piece of text,
or fill in the blanks in a sentence. This capability is used in many applications, from predictive
typing on smartphones to more complex tasks like coding assistance.

Code Generation: LLMs can be trained on code and thus generate code based on the users’ request.
Microsoft’s CoPilot is an example of using LLMs to generate codes.

3
DATA SHEE T | A Comprehensive Guide to Generative AI

GENERAL QUESTIONS ABOUT GENERATIVE AI

What are Generative AI, Foundation Models, and LLMs

Generative AI: Generative AI refers to a type of artificial intelligence that can create new content.
It’s designed to learn patterns in data and generate new content that fits the same patterns.
The most common use cases of generative AI include creating synthetic images, music, voiceovers,
and text. A popular form of generative AI is the generative adversarial network (GAN), which uses
two neural networks, a generator and a discriminator, to create and refine its output.

Foundation Model: Foundation models are a type of artificial intelligence model that is pre-trained
on a broad range of Internet text. They can understand and generate human language to a remarkable
degree. Examples of text foundation models are GPT-4 and PaLM. These models are referred
to as “foundation” because they can be fine-tuned for a wide variety of specific tasks, much like how
a foundation supports a range of structures above it.

LLM (Large Language Model): This term generally refers to a class of machine learning models that
learn to understand, generate, or translate human language. These models learn from large amounts
of text data and are capable of tasks such as answering questions, translating languages, summarizing
text, and more. Examples of such models are the Transformer-based models like BERT (Bidirectional
Encoder Representations from Transformers), GPT (Generative Pretrained Transformer), and others.

What are the capabilities of Foundation Models

Foundation Models, such as GPT-4 and DALL-E are versatile and capable of a wide range of tasks related
to understanding and generating human language. These models are trained on large corpora of text,
image, and audio data and can be fine-tuned for more specific tasks. Here are some examples
of the LLM capabilities:

Text Generation: These models are capable of generating human-like text. This can be used for marketing,
sales, or even for more practical purposes like drafting emails or other forms of communication.

Question Answering: Given a question, Foundation Models can generate an appropriate response
based on their understanding of the text. This is commonly used in customer service chatbots
and virtual assistants.

Translation: These models can translate text from one language to another. They have been trained on text
data in multiple languages and can thus understand and generate text in those languages.

Text Summarization: Foundation Models can take a long piece of text and create a shorter summary
that retains the most important points.

Information Extraction: Foundation models can extract structured information from unstructured
text data. For example, they can be used to extract names, dates, and places from a news article.

Completion and Prediction: Foundation Models can predict what comes next in a piece of text,
or fill in the blanks in a sentence. This capability is used in many applications, from predictive
typing on smartphones to more complex tasks like coding assistance.

Code Generation: LLMs can be trained on code and thus generate code based on the users’ request.

4
DATA SHEE T | A Comprehensive Guide to Generative AI

Microsoft’s CoPilot is an example of using LLMs to generate codes.

What are the different types of Foundation models

There are several types of foundation models, often categorized by the type of data they’re trained
on and the tasks they’re designed for:

Large Language Models (LLMs): Large language models, like GPT-3, BERT, and RoBERTa, are trained
on large text corpora and can perform a variety of natural language processing tasks, such as language
translation, summarization, question answering, sentiment analysis, and more.

Vision Models: These models, such as ResNet, EfficientNet, and Vision Transformer, are trained on large
image datasets and can perform tasks like object detection, image classification, and image generation.

Multimodal Models: These models are trained on a combination of different types of data, such as text
and images, and can perform tasks that require understanding multiple types of data. CLIP by OpenAI
is an example of a multimodal model that connects text and image understanding.

Reinforcement Learning Models: These models, like AlphaGo and OpenAI’s Dota 2 AI, learn by interacting
with an environment and receiving feedback. They can be used for tasks like game playing, robotics,
and other situations where an agent needs to make decisions based on its environment.

Language-Vision Models: Models like DALL-E and GPT-3.5-Turbo from OpenAI are examples of models that
combine language and vision capabilities, enabling tasks that involve generating images from textual
descriptions or understanding the content of images based on textual prompts.

Speech Models: These models, like Wav2Vec by Facebook AI, are trained on large amounts of audio data
and can perform tasks like speech recognition, speech generation, and audio classification.

Knowledge Graphs: Although not a machine learning model in the traditional sense, knowledge graphs
(like Google’s Knowledge Graph) are a type of foundation model that structures information into a network
of entities and their interrelations. They are used to enhance search results and enable more complex queries.

Each of these types of models represents a different approach to understanding and generating data,
and they each have their own strengths, weaknesses, and suitable application areas.

How Foundation Models are developed

Foundation models are developed through a two-step process: pre-training and fine-tuning.
Here’s a high-level overview of this process:

Pre-training: In this initial phase, the model is trained to predict or generate parts of the input data based
on the rest of the input. For instance, in the case of text-based models, they might be trained to predict
the next word in a sentence or to predict a masked word in a sentence. This pre-training is done on

a large corpus of text data sourced from the internet. During this process, the model learns to understand
the structure of the language, grammar, facts about the world, and even some reasoning abilities.

Fine-tuning: After pre-training, the model is fine-tuned on a more specific task. This involves training
the model on a smaller, task-specific dataset, where the model adjusts its parameters to get better at this
task. For example, a pre-trained model might be fine-tuned to answer medical questions, write Python code,
or generate poetry. The fine-tuning process allows the broad capabilities learned during pre-training
to be specialized to a specific task or domain.

5
DATA SHEE T | A Comprehensive Guide to Generative AI

The development of foundation models also involves significant computational resources, as these
models have millions or billions of parameters that need to be learned. The training process can take
several days or weeks, even on high-end hardware. Once the model is trained, it can generate responses
to prompts or answer questions by processing the input and generating the most likely output based
on what it learned during training.

How Foundation Models can be used in business

Foundation models, given their versatility and ability to understand and generate human-like text,
can be used in a variety of ways in the business context. Here are some potential applications:

Customer Support: Foundation models can be used to automate customer support, answering
common queries and providing information to customers. They can operate 24/7, handle multiple
queries simultaneously, and provide instant responses, improving efficiency and customer satisfaction.

Content Generation: These models can be used to generate content, such as marketing copy,
social media posts, or blog articles. They can help streamline the content creation process
and generate creative ideas.

Product Recommendations: Foundation models can be trained to understand user preferences

and behavior, and make personalized product recommendations.

Business Analytics: They can be used to analyze large amounts of text data, such as customer reviews
or feedback, to provide insights into customer sentiment and trends.

Email and Calendar Management: Foundation models can help manage emails and calendars,
drafting responses, scheduling meetings, and more, freeing up time for more strategic tasks.

Translation Services: Foundation models can provide instant translation services, helping businesses
communicate with customers or partners in different languages.

Data Extraction and Processing: Foundation models can be used to extract and process information
from unstructured data, such as documents or reports, saving time and reducing errors.

Training and Education: These models can be used to create interactive training materials,
provide explanations, or answer questions, enhancing learning experiences.

Sales and Marketing: Foundation models can be used to generate sales copy, personalize
marketing messages, or identify potential sales leads based on patterns in the data.

What are the strength of Foundation Models

Foundation models like LLMs have several strengths that make them powerful tools for a wide
range of applications:

Broad General Knowledge: These models are trained on a diverse range of internet text, which gives
them a broad base of general knowledge. They can generate outputs on a wide range of topics.

Language Understanding: Foundation models are very good at understanding and generating
human-like text. They can carry out sophisticated language tasks, such as translation, summarization,
question-answering, and more.

Contextual Understanding: These models can understand the context within a single conversation,

6
DATA SHEE T | A Comprehensive Guide to Generative AI

which allows them to generate coherent and contextually appropriate responses.

Versatility: The same model can be used for many different applications, from drafting emails to writing
code to creating poetry. This versatility makes them a powerful tool.

Scalability: These models can be used to automate tasks at scale. They can generate responses to thousands
of users simultaneously, making them useful for applications like customer service.

Creativity: Foundation models can generate creative outputs, like writing a story or a poem, making them
useful for creative applications.

Continual Learning: As new data becomes available, these models can be retrained to incorporate that data,
allowing them to continually improve and adapt to new information.

However, it’s important to remember that these strengths also come with limitations and risks. Careful
oversight and governance are needed to ensure that these models are used responsibly and ethically.

What are the weaknesses of Foundation Models

While Foundation Models offer many potential benefits, it’s important to remember that they also have
limitations and risks. They should be used as a tool to assist and augment human decision-making,
not to replace it. Furthermore, their use should comply with relevant laws and regulations, particularly
around data privacy and AI ethics. Businesses should also ensure that there are mechanisms in place
to review and verify the outputs of these models, to avoid errors or biases. Below are some of the limitations
regarding LLMs:

Inaccurate or Misleading Information: While these models have been trained on a vast corpus of information,
they can still output incorrect or misleading information. They generate text based on patterns they’ve
learned during training, but they don’t understand the information or verify its accuracy.

Lack of Contextual Understanding: While these models are good at understanding the context within
a single conversation, they struggle with understanding the larger context. For example, they don’t have
the ability to remember past interactions with users beyond the current conversation.

Knowledge Cutoff: These models have a “knowledge cutoff” beyond which they are not aware of new information.
For example, GPT-3’s knowledge is cut off in September 2021, and it doesn’t know about events or information
that came out after that date.

Dependency on Training Data: The performance of these models is heavily dependent on the quality
and diversity of the training data. If the data is biased or unrepresentative, the model’s outputs
will reflect those biases.

What are the risks and ethical concerns associated with the use of Foundation Models
Ethical Considerations: These models can potentially generate harmful or biased content, as they learn from
the data they were trained on, which may include harmful or biased text. They also do not understand ethics
or morality, and can’t make judgments about the appropriateness of their outputs.

Data Privacy: There are concerns about data privacy, as these models are trained on large amounts of data,
some of which may be personal or sensitive.

Misinformation and Disinformation: Given that these models generate text based on the patterns they’ve
learned during training, they can unintentionally spread misinformation or be used maliciously
to spread disinformation.

7
DATA SHEE T | A Comprehensive Guide to Generative AI

Harmful or Biased Content: Foundation models may generate content that is harmful, offensive,
or biased. They learn from the data they were trained on, which can include harmful or biased text.
Even with safeguards in place, it can be challenging to completely prevent such outputs.

Manipulation and Deception: There’s a risk that these models could be used to manipulate people
or deceive them, for example by generating fake news or deepfake content.

Economic and Social Impact: The widespread use of these models could have significant economic
and social impacts, such as job displacement due to automation, or changes in how we communicate
and interact.

Environmental Impact: The training of these models requires significant computational resources,
which has an environmental impact due to the energy usage.

Over-reliance: There is a risk of over-reliance on these models in decision-making processes, given their
limitations in understanding and context. This could lead to poor or erroneous decisions.

Lack of Explainability: The internal workings of these models are complex and not easily understood,
even by their creators. This lack of transparency can be a problem in applications where explainability
is important, such as in the legal or medical field.

DATAROBOT LLM INTEGRATIONS

How does DataRobot integrate with Azure OpenAI LLMs?

We integrate with Azure OpenAI in a number of ways. Utilizing LLMs to:

1. Improve efficiency / productivity of data scientists:

a. Notebook Code Assist
b. Insights Assist

2. Improve accuracy of automated machine learning blueprints by integrating Azure OpenAI LLMs
as custom tasks in DataRobot AutoML blueprints.

Does DataRobot share data with Azure OpenAI and Microsoft?

Yes, if you opt into the functionality DataRobot will send certain required user data to Azure OpenAI.
Data sent to Azure OpenAI is subject to Microsoft’s data processing policies, which can be found
at data, privacy and security documentation.

What kind of data is shared with Microsoft Azure for our Azure OpenAI integration?
Sensitive user data that range from user code for Notebook Code Assist to source data that was used
to build the models in DataRobot. More specifically, the data that will be shared with Microsoft
are the contents of the prompt which includes the contents of the current notebook.

Does DataRobot collect data from users?

DataRobot is collecting usage data (eg. user feedback and prompts) that will help us to improve our services
and prevent misuse. All collected data is subject to our privacy and data confidentiality terms. We store

8
DATA SHEE T | A Comprehensive Guide to Generative AI

usage data encrypted and we have an allow list who can access the usage data.

Will the Azure OpenAI integration be on by default for customers in preview mode,
or will that need to be turned on?
Notebook Code Assist is currently in Public Preview and needs to be turned on by request.

Will the Azure OpenAI integration be made available for Single-Tenant SaaS
or Self-Managed Installations?
The plans are still being developed and no decision is made at this time.

DATAROBOT GUIDELINE ON THE USE OF INTEGRATED LLM MODELS

Where can I find DataRobot guidelines on acceptable use of LLM integrated capabilities?
For Azure OpenAI integrations, please refer to Microsoft’s code of conduct on the appropriate
use and code of conduct of LLM integrated capabilities.

What is an acceptable prompt for DataRobot Notebook Code Assist?

For Azure OpenAI integrations, please refer to Mіcrosoft’s code of conduct on the appropriate use
and code of conduct of DataRobot Notebook Code Assist.

What is an acceptable prompt for the DataRobot Insight App?

This is not currently available in the product for customer use. Guidelines will be provided once
the product is available.

BEST PRACTICES ON THE USE OF LLMS

What are the best practices for prompt engineering?

Prompt engineering refers to the process of carefully crafting the input prompts that you give to
a machine learning model like GPT-3 to maximize the usefulness of the output it generates. This can be
a critical step in getting the most out of these models, as the way you phrase your prompt can significantly
influence the response.

Here are some best practices for prompt engineering:

Specificity: Make your prompts as specific as possible. Instead of asking “What’s the weather like?”,
ask “What’s the current temperature in San Francisco, California?” The latter is more likely to yield
the information you’re looking for.

Explicit Instructions: If you have a specific format or type of answer in mind, make that clear in your prompt.
For example, if you want a list, ask for a list. If you want a yes or no answer, ask for that.

9
DATA SHEE T | A Comprehensive Guide to Generative AI

Contextual Information: If relevant, provide some context to guide the model. For instance, if you’re asking
for advice on writing a scientific type of content, make sure to mention that in your prompt.

Use of Examples: When you want the model to generate in a particular style or format, giving an example
can help guide the output. For instance, if you want a rhyming couplet, you could include an example of one
in your prompt.

Prompt Length: While it can be useful to provide context, remember that longer prompts may lead the model
to focus more on the later parts of the prompt and disregard earlier information. Be concise and to the point.

Bias and Ethical Considerations: Be aware that the way you phrase your prompt can influence the output
in terms of bias and harmful response. Ensure your prompts are as neutral and fair as possible, and be aware
that the model can reflect biases present in its training data.

Temperature and Top-p Settings: Besides the prompt itself, you can also adjust the ‘temperature’
and ‘top-p’ settings. Higher temperature values make output more random, while lower values make it more
deterministic. The ‘top-p’ setting allows you to control the diversity of the output by limiting the model
to consider only a certain percentile of most likely next words.

Token Limitations: Be aware of the maximum token limitations of the model. For example, chatGPT 3 has
a maximum token limit of 4096 tokens. If your prompt is too long, it could limit the length of the
model’s response.

These are some of the key considerations in prompt engineering, but the exact approach can depend on the
specific use case, the model you’re using, and the kind of output you’re looking for. It’s often a process of trial
and error, and it can require a good understanding of both your problem domain and the capabilities
and limitations of the model.

Contact Us

225 Franklin Street, 13th Floor, Boston, MA 02110, USA

datarobot.com [email protected]

© 2023 DataRobot, Inc. All rights reserved.

DataRobot and the DataRobot logo are trademarks
of DataRobot, Inc.