There was an error in generating the vocab dictionary for the cleaned-up tokens. I have fixed the error in the new version 1.1.2 of babyGPT: [ #tokenization #GPT ].

Version 1.1.1 of babyGPT incorporates further improvements to the tokenizer training code. These upgrades get rid of superfluous tokens. A token is superfluous vis-a-vis a larger token if both predict exactly the same set of words in the training data.

Join millions who have switched to Grok.

Switch to this latest version of babyGPT (v. 1.1.0) for a better performing tokenizer. This version pre-filters the training data to get rid of long URL and other non-productive strings in media articles that should play no role in tokenizer training.

Just released Version 1.0.8 of babyGPT:.It is a PyTorch-Lightning adapted version of 1.0.7. The image shows example outputs during training. #LLM #chatbot #LanguageModeling

I have just finished the 2025 update of Computer and Network Security lecture notes:.

Although still a baby, babyGPT is growing and getting smarter: . Here are some prompts and their completions after 6 epochs of training:.#GPT #Chatbots

How is this for customer service: My dear friend Cris Bouman ran into problems with an on-line order for coffee beans. After a long and frustrating interaction with CS, she was finally told: "I'm going to loop in my team of real, live humans to help get this done for you.".

Wow! Java has just turned 30! .Java was the reason I fell in love with Object-Oriented.Programming. That led to my long Objects Trilogy Project with its 3 books:.You cannot be a world-class programmer if you have not mastered OOP.

With the inclusion of VQGAN, Version 2.5.5 of DLStudio incorporates all three foundational concepts in variational autoencoding: VAE, VQVAE and VQGAN: . Here are some additional thoughts regarding VQGANs:

The latest version of DLStudio has significantly improved implementations for the autoencoding classes: Autoencoder, VAE, and VQVAE. Now you can use a single parameter (num_repeats) for making arbitrarily deep the neural networks for these classes.

Was interesting to read this essay on the angst experienced by the academics as the development of BERT and then of GPT turned their world upside down. For those short on time, start with the beginning, switch to the end, and skip through the middle.

Here is a sample of the output of babyGPT after 8 epochs of training with a 85MB "athlete news" corpus. Trained with batch_size=200, embedding_size=128, max_seq_length=30 tokens, and num_attention_heads=4

If you like to play with #GPT models, here's is my babyGPT:. It's an end-to-end framework for collecting articles, for training a tokenizer, and for creating a GPT model you can play with in a typical university lab. Only Base Modelling for now.

I did not know Jeff Bezos was so eloquent. He said: "Freedom is ethical — it minimizes coercion — and practical — it drives creativity, invention, and prosperity." I am impressed!!!.

If you love jazz, you've got to see movie "For Love or Country". Although now old, the movie itself is timeless. Based on a true story, you'll love the dueling trumpets scene between Dizzie Guillespe (Charles Dutton) and Arturo Sandoval (pAndy Garcia).

I think you'd enjoy reading "Only fools think Elon is incompetent":.

CEOs who speak truthfully about things that impact their companies are rare. Hats off to Jensen Huang and Sam Altman for the courage with which they have spoken, the former about Quantum Computing and the latter about DeepSeek: .

Amazing! I am in awe!! Truly wonderful!!! For her and for Purdue!!!!!.Purdue University has a celebrity of its own!!!.

Truly amazing . Dick Van Dyke celebrating his 99th birthday with Coldplay. He still has the moves.

These two photos have got to be the most stunning I have ever seen. Just think that our existence (and survival) depends entirely on what's depicted in the photos!.