fix-ilql-sample-endings v. main

fix-ilql-sample-endings @20eec76/fix(ilql_randomwalks): bump `seq_length` to not trim any samples/2023-03-10 main @ded2e5e/fix(ppo_trainer): update `AdaptiveKLController` with correct KL (#361)/2023-03-13

Sorry

Created on March 13|Last edited on March 13

Comment

﻿
ppo_hh/pythia-6B-static-sft/7gpus﻿
Run set1
﻿
﻿
﻿
returns/mean
returns/mean
2004006008001k1.2k1.4kStep-1.5-1-0.50
losses/value_loss
losses/value_loss
2004006008001k1.2k1.4kStep0.10.20.30.40.50.6
returns/max
returns/max
2004006008001k1.2k1.4kStep-1.5-1-0.500.511.5
Run set1
﻿
ilql_sentiments/gpt2/1gpu﻿
Run set1
﻿
﻿
﻿
Run set1
﻿
sft_sentiments/gpt2/1gpu﻿
Run set1
﻿
﻿
﻿
Run set1
﻿
ppo_sentiments/gpt2-imdb/1gpu﻿
Run set1
﻿
﻿
﻿
Run set1
﻿
ilql_randomwalks/GPT2Config/1gpu﻿
Run set1
﻿
﻿
﻿
Run set1
﻿
ppo_randomwalks/randomwalks/1gpu﻿
Run set1
﻿
﻿
﻿
Run set1
﻿
﻿

Add a comment