This is a workbench where you can experiment with transformer LLM models. It aims at allowing a more "intimate" contact with GPT-like transformer models. What can transformers learn even without giga-data (and hopefully still fitting a GPU)?
It's through size that LLMs achieve amazing results, but size is a big barrier for normal people to play with them. The GPT-2 level models are on the border of what can be trained on commonly available GPUs and can thus be very valuable. They're not the smartest, but they are quite pliable.
GPTBench can be used to conveniently train a large or small transformer model and see what it can learn. It's made for tinkering and learning.
- Model sampling is simple and also includes a prompt mode where you can continuously interact with the model.
- You can train it starting from a blank model or from a pretrained GPT-2. Checkpoints can be loaded and saved at any point.
- Can measure accuracy, perplexity and loss metrics and can log training evaluations to .csv and TensorBoard formats.
- Usable in Python scripts or Jupyter notebooks or directly from the command line.
- Includes several Projects: training character-level and GPT2 token level datasets, toy models for learning to add two numbers, decimal-roman numeral translation, predicting number sequences. Also common tasks like prompting and fine-tuning GPT-2 level models.
- Includes Flash Attention and can train with gradient accumulation, so that larger batch sizes can be used in common GPUs.
This package grew from Andrej Karpathy's great minGPT, to whom I'm very grateful for the very inspiring lessons. The classes in model.py and trainer.py are mostly the same as in minGPT, adapted to work in this "workbench" and enhanced with some new features like Flash Attention and block_size cropping from nanoGPT.
Hoping this can be a contribution to a better understanding of these weird and fascinating machines, the decoding transformers.
Requires Python 3.7+ and PyTorch 2. Also uses NumPy, the tiktoken library (for the BPE tokenizer) and Hugging Face's transformers library (to download GPT-2 checkpoints).
You can run it in a plain CPU or CUDA GPU.
To use the gptbench package, download the repository and from the base directory (which has a setup.py script) do:
pip install -e .
Check the Projects, which include several examples, from toy data/models to prompting and fine-tuning pretrained GPT-2 models.
For examples, see Usage and the Jupyter notebooks and python scripts in Projects.
The way to select options in GPTBench is via settings in a config tree. See here for all the config options.
- Add examples of GPT2 fine tuning.
- Document dataset classes and how users can pass their own. As dataset classes become consolidated, derive from a base.
Good info about transformers and Large Language Models:
MIT