* 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
This commit is contained in:
Wing Lian
2024-03-18 13:10:00 -04:00
committed by GitHub
parent e8c8ea64b3
commit 2ea70ebbd8
14 changed files with 451 additions and 24 deletions

View File

@@ -191,6 +191,11 @@ def normalize_cfg_datasets(cfg):
f"updating dataset {ds_cfg.path} with `conversation: chatml` to match your chat_template"
)
cfg.datasets[idx].conversation = "chatml"
if ds_cfg.type == "orpo.chat_template" and not ds_cfg.chat_template:
LOG.info(
f"updating dataset {ds_cfg.path} with `chat_template: chatml` to match your chat_template"
)
cfg.datasets[idx].chat_template = "chatml"
def validate_config(cfg: DictDefault, capabilities: Optional[dict] = None):