askai VS nanoGPT

A while back, I wrote a simple CLI-wrapper around OpenAI's API that I'm using daily (https://github.com/maxvfischer/askai). I use it as an addition to Stackoverflow to ask quick programming-related question straight in the terminal.

Nice work! I built something similar years ago and I did compile the probabilities based on a corpus of text (public domain books) in an attempt to produce writing in the style of various authors. The results were actually quite similar to the output of nanoGPT[0]. It was very unoptimized and everything was kept in memory. I also knew nothing about embeddings at the time and only a little about NLP techniques that would certainly have helped. Using a graph database would have probably been better than the datastructure I came up with at the time. You should look into stuff like Datalog, Tries[1], and N-Triples[2] for more inspiration.

You're idea of splitting the probabilities based on whether you're starting the sentence or finishing it is interesting but you might be able to benefit from an approach that creates a "window" of text you can use for lookup, using an LCS[3] algorithm could do that. There's probably a lot of optimization you could do based on the probabilities of different sequences, I think this was the fundamental thing I was exploring in my project.

Seeing this has inspired me further to consider working on that project again at some point.

[0] https://github.com/karpathy/nanoGPT

[1] https://en.wikipedia.org/wiki/Trie

[2] https://en.wikipedia.org/wiki/N-Triples

[3] https://en.wikipedia.org/wiki/Longest_common_subsequence

where x1 denotes the 1st token, x2 denotes the 2nd token and so on, respectively.

I understand the conditional terms p(x_n|...) where we use cross-entropy to calculate their losses. However, I'm unsure about the probability of the very first token p(x1). How is it calculated? Is it in some configurations of the training process, or in the model architecture, or in the loss function?

IMHO, if the model doesn't learn p(x1) properly, the entire formula for Bayes' rule cannot be completed, and we can't refer to LLMs as "truly generative". Am I missing something here?

I asked the same question on nanoGPT repo: https://github.com/karpathy/nanoGPT/issues/432, but I haven't found the answer I'm looking for yet. Could someone please enlighten me.

This repo, albeit "old" in regards to how much progress there's been in LLMs, has great simple tutorials right there eg. fine-tuning GPT2 with Shakespeare: https://github.com/karpathy/nanoGPT

For training from scratch, maybe a small model like https://github.com/karpathy/nanoGPT or tinyllama. Perhaps with quantization.

It does remind me of a project [1] Andrej Karpathy did, writing a neural network and training code in ~600 lines (although networks have easier logic to code than a compiler).

[1] https://github.com/karpathy/nanoGPT

But I'm still not convinced that it can't in theory. Maybe the training set or transformer size I'm using is too small. I'm using nanoGPT implementation (https://github.com/karpathy/nanoGPT) with layers 24, heads 12, and embeddings per head 32. I'm using character-based vocab: every digit is a separate token, +, = and EOL.

https://github.com/karpathy/nanoGPT/blob/f08abb45bd2285627d1...

At training time, probabilities for the next token are computed for each position, so if we feed in a sequence of n tokens, we basically get n training examples, one for each position, but at inference time, we only compute the next token since we’ve already output the preceding ones.

And there are a bunch of other efforts at making training more efficient. Here's a cool model by Karpathy (OpenAI/used to head up Tesla's efforts): https://github.com/karpathy/nanoGPT

Just being a part of any auto-regressive system does not contradict his statement.

Go look at the GPT training code, here is the exact line: https://github.com/karpathy/nanoGPT/blob/master/train.py#L12...

The model is only trained to predict the next token. The training regime is purely next-token prediction. There is no loopiness whatsoever here, strange or ordinary.

Just because you take that feedforward neural network and wrap it in a loop to feed it its own output does not change the architecture of the neural net itself. The neural network was trained in one direction and runs in one direction. Hofstadter is surprised that such an architecture yields something that looks like intelligence.

He specifically used the correct term "feedforward" to constrast with recurrent neural networks, which GPT is not: https://en.wikipedia.org/wiki/Feedforward_neural_network

Does anyone have or know of an example implementation in plain pytorch, not huggingface transformers. Like something you could plug into https://github.com/karpathy/nanoGPT ?