Montinger / Transformer-Workbench

Playground for Transformers
42 stars 16 forks source link

Transformer-Workbench

Playground for some experiments with Transformer models.

transformer-from-scratch

This folder contains code, which implements a transformer from scratch and trains it on the original Attention is all you need translation task. See the README in this folder for further details.

DitBERT

This folder contains the code to define and train my own BERT-like model DitBERT, which is also benchmarked with GLUE against BERT-base and RoBERTa-base. See the README in this folder for further details.

parse-text-corpora

Contains the code used to parse the text corpora, used to train e.g. DitBERT.

LoRA-from-scratch

A from scratch implementation of Low-Rank Adaptation (LoRA) for a RoBERTa model. This blogpost also shows an optimized version of QLoRA (Quantized LoRA) with the huggingface PEFT library and bitsandbytes for quantization.