Imaginary transcripts of Red Bull Music Academy (RBMA) lectures, generated using GPT-2 via gpt-2-simple. The medium (355M parameter) model was fine-tuned using actual RBMA lectures (~500 transcripts, ~20MB of text) for 1,000 steps.
There are a selection of dumps (lol) of transcripts for each "temperature" - the higher the temperature, the weirder (and less syntactically correct) the lectures are.
temp_0_7: Normal and syntactically correcttemp_0_7_bellerin: Normal and syntactically correct but about Héctor Belleríntemp_1_0: Wilder and less syntactically correcttemp_1_3: Very extremely wild O_o
The actual transcripts are available in plain-text format within the rbma-lectures repo, taken from the RBMA lectures mini-site on 3rd November, 2019.
The files within the data/ subdirectory of the rbma-lectures repo were appended into a single .txt file and used to retrain the model with gpt-2-simple using a GPU via a copy of this Colaboratory notebook. (I haven't shared the retrained model because it is very large). The same notebook was used to generate the new transcripts and then downloaded locally.