2ea70ebbd8
* orpo trainer * rl handling for orpo * support for remove_unused_columns * orpo fixes * fix loader for orpo * chore: lint * fix default for remove_unused_columns * roll ORPO into the main AxolotlTrainer so it can be compatible with some of the other techniques like relora * better handling of system message for orpo * revert system prompt changes for chat templtes * no need for else condition * split dataset parsing into it's own component
1.6 KiB
1.6 KiB
RLHF (Beta)
Overview
Reinforcement Learning from Human Feedback is a method whereby a language model is optimized from data using human feedback. Various methods include, but not limited to:
- Proximal Policy Optimization (PPO) (not yet supported in axolotl)
- Direct Preference Optimization (DPO)
- Identity Preference Optimization (IPO)
RLHF using Axolotl
Important
This is a BETA feature and many features are not fully implemented. You are encouraged to open new PRs to improve the integration and functionality.
The various RL training methods are implemented in trl and wrapped via axolotl. Below are various examples with how you can use various preference datasets to train models that use ChatML
DPO
rl: dpo
datasets:
- path: Intel/orca_dpo_pairs
split: train
type: chatml.intel
- path: argilla/ultrafeedback-binarized-preferences
split: train
type: chatml.argilla
IPO
rl: ipo
ORPO
Paper: https://arxiv.org/abs/2403.07691
rl: orpo
orpo_alpha: 0.1
remove_unused_columns: false
chat_template: chatml
datasets:
- path: argilla/ultrafeedback-binarized-preferences-cleaned
type: orpo.chat_template
Using local dataset files
datasets:
- ds_type: json
data_files:
- orca_rlhf.jsonl
split: train
type: chatml.intel
Trl autounwrap for peft
Trl supports autounwrapping peft models, so that a ref model does not need to be additionally loaded, leading to less VRAM needed. This is on by default. To turn it off, pass the following config.
# load ref model when adapter training.
rl_adapter_ref_model: true