Though the overwhelming majority of our explanations rating poorly, we imagine we are able to now use ML methods to additional enhance our capacity to provide explanations. For instance, we discovered we had been capable of enhance scores by:
- Iterating on explanations. We will improve scores by asking GPT-4 to give you doable counterexamples, then revising explanations in mild of their activations.
- Utilizing bigger fashions to offer explanations. The typical rating goes up because the explainer mannequin’s capabilities improve. Nonetheless, even GPT-4 provides worse explanations than people, suggesting room for enchancment.
- Altering the structure of the defined mannequin. Coaching fashions with completely different activation capabilities improved rationalization scores.
We’re open-sourcing our datasets and visualization instruments for GPT-4-written explanations of all 307,200 neurons in GPT-2, in addition to code for rationalization and scoring utilizing publicly accessible fashions on the OpenAI API. We hope the analysis group will develop new methods for producing higher-scoring explanations and higher instruments for exploring GPT-2 utilizing explanations.
We discovered over 1,000 neurons with explanations that scored not less than 0.8, that means that in line with GPT-4 they account for a lot of the neuron’s top-activating habits. Most of those well-explained neurons will not be very fascinating. Nonetheless, we additionally discovered many fascinating neurons that GPT-4 did not perceive. We hope as explanations enhance we could possibly quickly uncover fascinating qualitative understanding of mannequin computations.