This repository has been archived by the owner on Feb 25, 2022. It is now read-only.
-
Notifications
You must be signed in to change notification settings - Fork 949
Pull requests: EleutherAI/gpt-neo
Author
Label
Projects
Milestones
Reviews
Assignee
Sort
Pull requests list
Nuke stitch; simplify config; sh{ard,uffle} files
#28
by shawwn
was closed Nov 19, 2020
Loading…
updated Nov 19, 2020
improve on causal linear attention
#68
by lucidrains
was merged Nov 1, 2020
Loading…
updated Nov 1, 2020
add faster sampling for global attention
#43
by lucidrains
was merged Sep 12, 2020
Loading…
updated Oct 10, 2020
3 of 4 tasks
logic for automatically generating dataset config file
#59
by lucidrains
was merged Sep 22, 2020
Loading…
updated Oct 10, 2020
make dataset shuffling and interleaving deterministic by addition of …
#63
by lucidrains
was merged Sep 22, 2020
Loading…
updated Oct 10, 2020
make sure datasets skips the number of batches equal to the current g…
#62
by lucidrains
was merged Sep 22, 2020
Loading…
updated Oct 10, 2020
Print warning if truncating input
#64
by sdtblck
was merged Sep 23, 2020
Loading…
updated Sep 23, 2020
Pw/resume deterministic dataset correctly
#61
by lucidrains
was closed Sep 20, 2020
Loading…
updated Sep 20, 2020
move scripts for tokenization and creating tfrecords to root, also ad…
#58
by lucidrains
was merged Sep 19, 2020
Loading…
updated Sep 19, 2020
training with multiple GPUs, update readme
#56
by lucidrains
was merged Sep 19, 2020
Loading…
updated Sep 19, 2020
allow use of tokenizer from gpt2 when creating tfrecords
#57
by lucidrains
was merged Sep 19, 2020
Loading…
updated Sep 19, 2020
add masked language modeling, for training BERT or RoBERTa like atten…
#47
by lucidrains
was merged Sep 13, 2020
Loading…
updated Sep 13, 2020
do not count padding tokens in loss
#45
by lucidrains
was merged Sep 13, 2020
Loading…
updated Sep 13, 2020
fix big bug with exclamation mark being used as padding for gpt2
#50
by lucidrains
was merged Sep 13, 2020
Loading…
updated Sep 13, 2020
ProTip!
Follow long discussions with comments:>50.