Predictive Data Debugging: See and Control What Your Model Will Learn Before Training

A model's performance is determined by its dataset, and data is essentially how you program a model. However, the existing DPO (Direct Preference Optimization) approach has the inconvenience of only being able to trace back what went wrong after training results are already in.

The new Predictive Data Debugging technique predicts, before training even begins, what behaviors a given dataset will reinforce or suppress in the model. This predictive model shows an extremely high correlation of R² = 0.9 with actual training outcomes.

The core of this technique is interpreting data by leveraging the model's interpretability. Instead of simply using embedding-based clustering, it directly measures the concepts the model computes while processing data, allowing precise identification of what the model will learn.

This provides the following benefits:

• Prevention: Datasets can be corrected before unwanted behaviors (e.g., sycophancy, hallucination, inappropriate answers) emerge.
• Precise tracing: You can pinpoint exactly which data clusters trigger specific behaviors.
• Intentional model design: You can directly control the model's training signal by restructuring the dataset and training process.

Goodfire has integrated this technology into its platform, Silico, to help developers design models exactly as they intend.