Using Reinforcement Learning to Guide Chains of Thought

• Uses Hugging Face TRL for PPO
• Uses Hugging Face Peft for LoRA.
• Uses Bitsandbytes internally for 4bits and 8bits reference model modes.
• Uses our QLora standalone lib for QLora.

Sets of TRL and/or SFT jobs:

Launch jobs with

./job_sets/launch_sets.py <job_set_name>   

Check the status with:

./job_sets/check_status.py

With TRL:

Where the reinforcement learning is located.

./with_trl/launch.py <experiment_name>

Approach SFT:

./approach_sft/launch.py <experiment_name>