A small GPT model based on a character level tokenization system. It can train on any dataset and learns patterns really well. Adjust as needed.
To Train/Fine-Tune:
- Add an input.txt file with any text you want it to learn from. (
<eos>will be converted into an end of statement token) - Run train.py to train it. (Inside it you can change batch_size, max_iters, learning_rate to march your needs)
- NOTE: If you have a model.pth file already, it will be loaded from at the start of training and overwritten at the end of training.
To prompt: Just run prompt.py.
NOTE: This is an output from model.pth which was trained on a cleaned simple wikipedia custom dataset, and a Q&A/Basic Conversation Dataset custom dataset for 30,000 max_iters each, then fine-tuned on a few geography trivia questions. The training was done on google collab with the fine-tuning being done on my MacBook.
Prompt: Hello!
Output: Hey!
Prompt: What's your name?
Output: TinyTalker.
Prompt: What can you do?
Output: I can generate text!
Prompt: Nice!
Output: Cool!
Prompt: What is the capital of Japan?
Output: Tokyo is the capital of Japan.
Prompt: What are some other cities there?
Output: Sure! Here is a list of Japan: Tokyo(Japan's Capital), Osaka, Kyoto, Sapporo, Sendai, Kobe, Fukuoka, Nagoya, and Nara are some Japanese cities.
Prompt: Can you give me even more Japanese Cities?
Output: Shi Hiroshima and Nagasaki famous for being the only 2 cities to be Nuked during World War 2.
Prompt: Wow... thats grim.
Output: What cities were nuked during World War 2/WWWWWW2?
Prompt: You just told me.
Output: TinyTalker.
Prompt: Ah okay, this is roughly the limit of what you can do.
Output: Nagasaki as the gateway to the Amazon.
Prompt: yep.
Output: Manaus is situated in which country?
Prompt: Brazil, anyways Bye! Have a nice day!
Output: Bye!
This model is inspired by Andrej Karpathy Let's build GPT from scratch video aswell as Andrej Kaparthy nanoGPT with modifications for this project.