Description:
The current configuration system relies on separate YAML files for each training method (SFT, DPO, RL). While this is functional, it can lead to code duplication and inconsistencies. This issue proposes the implementation of a unified configuration system that centralizes the management of all training parameters.
Tasks:
Create a config.py module to define a base configuration class.
Implement data validation to ensure that all required parameters are present and have the correct type.
Refactor the training scripts and notebooks to use the new configuration system.
Update the documentation to reflect the changes.
Description:
The current configuration system relies on separate YAML files for each training method (SFT, DPO, RL). While this is functional, it can lead to code duplication and inconsistencies. This issue proposes the implementation of a unified configuration system that centralizes the management of all training parameters.
Tasks:
Create a config.py module to define a base configuration class.
Implement data validation to ensure that all required parameters are present and have the correct type.
Refactor the training scripts and notebooks to use the new configuration system.
Update the documentation to reflect the changes.