Since it is a very large project containing a huge variety of tasks related to the structured knowledge, the code unification becomes a unique challenge.

We specify hyperparameters covered by HuggingFace's training arguments in the command line, while we use config files to specify tasks and hyperparameters not covered by HuggingFace's training arguments.

Up to now, we divide all the config files into two groups: META_TUNING and Salesforce.

Each file under META_TUNING/ corresponds to a task with a particular setting. You can add a config file for a new task or for an existing task with a new setting. You can also add new task-specific hyperparameters into an existing config file if you want to use them elsewhere in the code.

Each file under Salesforce/ corresponds to an experiment, which specifies 1) files (child configs) under META_TUNING/ and 2) experiment-specific hyperparameters. You can add a config file for a new experiment. You can also add new experiment-specific hyperparameters into an existing config file if you want to use them elsewhere in the code.