Proposed synchronisation of RL and LM approaches#4

Open

thomfoster wants to merge 24 commits intoCarperAI:mainfrom

thomfoster:main

Commits on Dec 5, 2022

collect rollouts for hyperparameter sweep on roc_story
Thomas Foster
committed
update reward model to one with wider range of sentiment options
Thomas Foster
committed

Commits on Jan 5, 2023

update for relative logging paths
Thomas Foster
committed
add requirements.txt
Thomas Foster
committed
add incredibly overengineered script for decoding rollouts
Thomas Foster
committed
rename decode_rollout > decode_rollouts
Thomas Foster
committed
rename ppo_roc_story_sentiment_rollouts.py >> generate_ppo_roc_story_sentiment_rollouts.py to reflect that this is the script to generate data, not the class that uses it
Thomas Foster
committed
add class to use rollout data as language modelling task
Thomas Foster
committed
rename to reflect that this is a generic wrapper for any rollout, not just roc stories
Thomas Foster
committed
super simple script to train lm with accelerate on the rollout data
Thomas Foster
committed

Commits on Jan 6, 2023

add eval loop
Thomas Foster
committed
add verbosity flag for dataset
Thomas Foster
committed
use verbosity flag
Thomas Foster
committed
add flag for yielding prompt only for generation during evaluation
Thomas Foster
committed
fix iterator
Thomas Foster
committed
add wandb logging
Thomas Foster
committed
add dataset shuffling
Thomas Foster
committed

Commits on Jan 29, 2023

folder level refactor
Thomas Foster
committed
add mega simple example script
Thomas Foster
committed
fix train script
Thomas Foster
committed
ipdate readme
Thomas Foster
committed
Merge branch 'main' into task/synchronize-rl-and-lm-approaches
thomfoster
authored
Merge pull request #1 from thomfoster/task/synchronize-rl-and-lm-approaches
thomfoster
authored
seperate into rl and lm
Thomas Foster
committed