Curious on LLM from Scratch - Chapter 5 Training Example

 Running the Chapter 5 Training Example, 

This is an interesting example and without fully understanding it, asking Github Copilot, Chat GPT abotu the code.

...


Here it is running, and also updates on the code.



 The model has approximately 124 million parameters (hence the name) primarily from:

 Position Embeddings: context_length × emb_dim = 256 × 768 parameters

 Transformer Layers: 12 layers, each with attention heads and feed-forward networks


 Even though "The Verdict" is only 3,600 words:

 The model is designed to learn general language patterns, not just memorize the text

  The vocabulary size (50,257 tokens) matches GPT-2's full vocabulary

 The architecture follows smaller GPT-2 specifications (768-dim embeddings, 12 layers)

 This is the total number of unique tokens the model can recognize and generate

'''

'''

The full vocabulary (50,257 tokens) gives it the capacity to represent 

many words it never sees in training

It's based on the GPT-2 design which was trained on millions of documents

'''

'''

The word "Every" is converted to its corresponding token ID from the GPT-2 vocabulary.

Embedding Layer: The model has an embedding layer for all 50,257 tokens, including "Every".

See:

https://github.com/rasbt/LLMs-from-scratch/tree/main/ch05

And:

https://github.com/berlinbrown/berlin-learn-ml-dl-capstone-projects/tree/main/basic-exercises/basic-rastb-fork-chap5

Comments

Post a Comment

Popular posts from this blog

JVM Notebook: Basic Clojure, Java and JVM Language performance

Image
"Measure, don't guess. The primary goal of all performance tuning exercises should be maximize the end user experience given the resource constraints." [1] "It isn't so much a "farewell to the J" as an expansion of the platform opportunities Java provides. Sun's investment to power ongoing development of JRuby and Jython broadens the range and reach of Java, as a whole." -- Rick Ross Overview and JVM Languages    One of the exciting trends to recently emerge from the Java community is the concept of the JVM language. These technologies are all that you would expect them to be. They are implementations of languages that run on the Java Virtual Machine. Some are newly created and some are based on existing, more mature languages. JRuby, Jython are two JVM languages based on CRuby and CPython. Groovy, Scala, Clojure are three completely new JVM languages that were created to add new language features that weren't supported by the core Ja...
Read more

On Unit Testing, Java TDD for developers to write

It has been a while, let's kick off 2017 with a blog entry. I have read and I am reading about four or five posts a day about unit testing. It really has been a long time obsession for me. I have moved past the technical and practical considerations on unit testing frameworks and done with the debates with "should you use Junit or Mockito or Karma?" I am more interested in the psychology of unit testing, who does it, likes it, hates it? It really is one of those easy to learn, hard to master concepts. For example, many many may play chess when they are young and can end up being horribly chess players most of their life, I am part of that majority. Unfortunately, I have never played chess and sat down for hours and tried to master it. I never see the common patterns or have a developed end game. I mostly just play with a knowledge of the basic rules. Following good unit testing practices within your software development shop is a lot like playing chess. It is easy to le...
Read more

Application server performance testing, includes Django, ErlyWeb, Rails and others

Image
The following represents a set of toy benchmark results for various web application servers including ErlyWeb, Hunchentoot, Django and other application servers. The goal was to test out of the box performance through a simple VIEW page. For example, I was trying to avoid static files but going through the controller and the view. In the Rails example, there is a simple controller that then activates the view template. The same approach is used with the ErlyWeb example. Ideally, we are trying to avoid the server loading the static files but in our case the server may be loading cached content. Overall, Erlang's ErlyWeb/Yaws had great performance and out of the box Django did very well. Tomcat and my various low-level simple JSP/Servlet apps also had stable performance. Software and Hardware Configuration used in the Benchmark Two machines were used in the benchmark, a Win32 dual core machine and a single CPU Linux machine. Win32 Box: Microsoft Windows XP, Vers 2002, Service ...
Read more