Generative AI

Introduction to Generative AI, LLMs and harnessing the power of LLMs for improving productivity and personal growth
 Goal for today

Understanding the basics of Learn how to leverage the Engineering our existing Harnessing the power of LLM
LLMs power of LLMs to make our application's code in daily activities.
lives easier for free.
Introduction
• LLM deep learning models are Understands & generates
contextually aware text. In the simplest of terms, LLMs are next-
word prediction engines.
• GPT – GPT models are neural network-based language prediction
models built on the Transformer architecture. They analyze
natural language queries, known as prompts, and predict the best
possible response based on their understanding of language
• Transformers : Deep neural networks.
How LLMs works?
Encoded Decoded
Input text LLM Model Output
Embedding (vector) Embeddings (text)
How language model works under the hood
How language
model works
under the hood
(Cont.)
 • Massive Multitask Language Understanding – Datasets
o MMLU (Medical Genetics),MMLU (College Biology),MMLU (Professional
medicine),MMLU (Econometrics),MMLU (High School Computer
Science),MMLU (College Chemistry),MMLU (High School Mathematics),MMLU
(Electrical Engineer),MMLU (College Physics),MMLU (High School
Statistics),MMLU (College Mathematics),MMLU (Abstract Algebra),MMLU
(Astronomy),MMLU (Elementary Mathematics),MMLU (Formal Logic),MMLU

LLMs (High School Biology),MMLU (Machine Learning),MMLU (College Computer

Science),MMLU (High School Physics),MMLU (High School Chemistry)..

Datasets & • HellaSwag – Dataset

benchmarks o Logical reasoning tasks which need commonsense like humans

• ARC Dataset
o AI2 Reasoning Challenge (contain 14.5+ million multipe choice questions which
are difficult and answering them need human like knowledge

• WinoGrad – Dataset
o For commonsense detection, this models contains complex dataset
Terms used in LLM
1. Instruction / Prompt - a specific task or instruction you want to achieve using the LLM model
2. Context - external information or additional context that can steer the model to better responses
3. One-shot query – By providing some context you can ask a question
Example: Translate the following English sentences into French. Here's an example: 'I love playing
soccer.
4. Zero-shot query – Directly asking a question
Example : What is the capital city of India
5. Few-shot query - By providing all the information you can ask questions
Example: "Here is my blood test report pls pull out hemoglobin and RBC values and WBC values along
with observed values as a json object with key and values
In this format {"observation_name": "hemoglobin", "value":"15.6"}
Hemoglobin 15.6 mg/dl 12 – 16,
RBC 2.7 lacs 1.5– 3.5
WBC 5200 thou 4000 – 6000 "
 Most popular LLMs like ChatGPT, Google
Bard

o Can summarise text

o Can generate APIs and code
Capabilities o Can fix the bugs in code
of most o Can rewrite the algorithms in existing code
o Can generate formal mail
popular o Can rewrite / paraphrase context
o Can help us plan milestone and timelines of
LLMs projects
o Can brainstorm and provide ideas
o Can provide us the travel plans
o Can troubleshoot anything in any domain by
providing proper information to the context
Can summarise text

Input prompt

Output
Can generate APIs and code

Input prompt

Output
Can generate formal mail
Output

Input prompt
Can rewrite / paraphrase context

Input prompt Output

Can help us plan milestone and timelines of projects
Prompt

Output
Can brainstorm and provide ideas
Prompt

Response
Can provide us the travel plans
Prompt

Response
Can troubleshoot anything in any domain by providing proper information
to the context
LLM based document search use case
Media specific models