Structured environments for training data

We study what a training example needs to be understood and evaluated consistently. Rules define valid actions and successful outcomes. Context explains the request and response.

We will check whether examples include the needed information and whether reviewers evaluate them consistently. We will then test whether these qualities help a model use context and follow the task.

Simulation games give players defined tasks and rules. In the packing example below, players fit fixed food shapes into a bento. The rules define valid moves and when the board is complete.

Argotu is a social platform. We study how to preserve the context, intention, and perspective of its exchanges when turning them into training examples.

We will test data quality and model performance separately. For the data, we will check whether examples are clear, consistent, keep the needed context, and cover the task. For the model, we will check whether responses follow the request, use context, and respond to clarification.

We will also test tasks outside Argotu and simulation games to see whether models can use context in unfamiliar situations.