Text Represantations:LLMs

How does LLMs work?

Learning Network Training process in simple terms include Big Data input and output as response

• Large Language Models are simply trying to Predict the next word in the sentence based from what they learned or trained on
  For Example Normally : Peanut butter and ? (click your answer below)
  1.Meat
  2.Jelly
  3.Jam

Types:Bidirectional predict the previous and next word, and Unidirectional predict the next word only

Capabilities:

• They produce text
• They can retrieve some information
• Maintain Dialogue
• Sequential
• Language fluency they can manupilate language easily
• Good for Traslating languages

---

Limitation:

• They cannot plan
• Hallucinations *Way of Expressing Error
• They can fairly Reason
• Too expensive to train
• They just generate the next token without planning the answer
• They are not 100% factual
• Its Mixbag when it comes to Learning with Experience

---

Facts Check

• They lack understanding of how the world works.No common sense
• Not Sentiment.
• They are Predictive Algorithms,predicting the next token.
• They are Probalistic.
• Deep Fake
• Generative and Relative models.
• They can discover staff but cannot create new things

---

Generative PreTrained Transformers

Models:GPT 1-4 Plus.

Type : Generative Dialogue Systems or Chatbots.
Self Supervised, Auto Regressive LLMs ChatGpt



What Tasks are good for GPTs

GPT Applications (Unidirectional Transformers) predicting the next word only

Brain Storm: creative stories, poetry, and personalized social media content.

Text Generation and Content Creation: Automated writing for blogs, creative stories, poetry, and personalized social media content.

Conversational Agents: Chatbots for customer service, virtual assistants for scheduling, and dialogue generation in gaming.

Programming Assistance: Tools like GitHub Copilot for code generation, helping developers write and debug code more efficiently.

Personalization and Recommendations: Content recommendation, dynamic ads, and personalized email campaigns.

Education Tools: AI-driven tutoring systems, essay grading, and real-time feedback for students.

Entertainment and Gaming: Interactive storytelling in video games, and virtual characters with dynamic dialogue generation.

---




GPT Model Evolution Bench Marks

Version	Year Released	Accuracy (%)	Task Performance (Steamrolled Tasks)
GPT-1	2018	75%	Poor – limited ability in multi-step reasoning
GPT-2	2019	80%	Moderate – some capability in steamrolling tasks, but inconsistent
GPT-3	2020	85%	Good – can handle basic multi-step reasoning
GPT-3.5	2021	87%	Improved – handles steamrolled tasks with greater accuracy and depth
GPT-4	2023	90%	Excellent – demonstrates strong capability in steamrolling complex tasks with multi-step reasoning

BERT

Bert Learn More

---

Evolution of Bidirectional Transformers

Model	Year Released	Key Features	Common Tasks
BERT	2018	Bidirectional context, Masked Language Model, Next Sentence Prediction	Sentiment Analysis, Question Answering, NER, Text Classification
RoBERTa	2019	Improved BERT with more training data, No Next Sentence Prediction	Text Classification, Sentence Pair Tasks, NER
ALBERT	2019	Parameter Sharing, Reduced Memory Load, Faster Training	Question Answering, Sentence Classification, NER
DistilBERT	2019	Distilled BERT, Faster Inference, 60% Fewer Parameters	Text Classification, Sentence Pair Tasks, NER
T5	2019	Text-to-Text Framework, Encoder-Decoder, Versatile NLP Model	Translation, Text Generation, Summarization, Question Answering

Real-World Applications of GPT and BERT

Visual Represantations