This PR adds the create_merge_dataset.py and train_dam.py scripts by adding new command line arguments to support additional functionality.
New Argument:
--base_model_dataset_name: Allows users to specify a sample dataset name related to the base model. This
Summary
This PR adds the create_merge_dataset.py and train_dam.py scripts by adding new command line arguments to support additional functionality.
Changes in create_merge_dataset.py
New Argument:
--base_model_dataset_name: Allows users to specify a sample dataset name related to the base model. This functionality is to use a dataset similar to the base model's training data, with the default name set to reflex-ai/fineweb-ultra-mini.
Changes in train_dam.py
New Argument:
--loss_base_data_dist: Computes the distribution difference between the base model and the merged model for the base data.
Summary
This PR adds the
create_merge_dataset.py
andtrain_dam.py
scripts by adding new command line arguments to support additional functionality.New Argument:
--base_model_dataset_name
: Allows users to specify a sample dataset name related to the base model. ThisSummary
This PR adds the
create_merge_dataset.py
andtrain_dam.py
scripts by adding new command line arguments to support additional functionality.Changes in
create_merge_dataset.py
--base_model_dataset_name
: Allows users to specify a sample dataset name related to the base model. This functionality is to use a dataset similar to the base model's training data, with the default name set toreflex-ai/fineweb-ultra-mini
.Changes in
train_dam.py
--loss_base_data_dist
: Computes the distribution difference between the base model and the merged model for the base data.