Skip to content

Latest commit

 

History

History
23 lines (12 loc) · 490 Bytes

File metadata and controls

23 lines (12 loc) · 490 Bytes

transformer

Andrej Karpathy follow along

Objective

Recreate the decoder-only transformer from bottom up

Generate coherent WikiHow articles

Data

Wikihow corpus

Results

Sub par text generation results because of compute constraints (my potato laptop)

future improvements

Use GPU

Use BPE, Wordpiece etc for tokenization. The character level tokenization method is simplistic and fails to capture statistics of the corpus