GitHub - PiotrNawrot/nanoT5: Fast & Simple repository for pre-training and fine-tuning T5-style models en (github.com)

This repository comprises the code to reproduce the pre-training of a "Large Language Model" (T5) under a limited budget (1xA100 GPU, < 24 hours) in PyTorch. We start from the randomly initialised T5-base-v1.1 (248M parameters) model, and we pre-train it on the English subset of the C4 dataset and then fine-tune it on Super-Natural Instructions (SNI) benchmark.

In ~16 hours on a single GPU, we achieve 40.7 RougeL on the SNI test set, compared to 40.9 RougeL of the original model weights available on HuggingFace Hub and pretrained on 150x more data through "a combination of model and data parallelism [...] on slices of Cloud TPU Pods", each with 1024 TPUs.

Our core contribution is not the T5 model itself, which follows the HuggingFace implementation. Instead, we optimise everything else in the training pipeline to offer you a user-friendly starting template for your NLP application/research. Most importantly, we show that it is possible to pre-train the T5 model to the top performance under a limited budget in PyTorch.