Though the overwhelming majority of our explanations rating poorly, we imagine we will now use ML methods to additional enhance our capability to supply explanations. For instance, we discovered we have been in a position to enhance scores by:
- Iterating on explanations. We are able to improve scores by asking GPT-4 to provide you with potential 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. Nevertheless, even GPT-4 offers worse explanations than people, suggesting room for enchancment.
- Altering the structure of the defined mannequin. Coaching fashions with totally different activation features 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 using publicly available models on the OpenAI API. We hope the analysis neighborhood 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 at the least 0.8, which means that in accordance with GPT-4 they account for many of the neuron’s top-activating habits. Most of those well-explained neurons usually are not very fascinating. Nevertheless, we additionally discovered many fascinating neurons that GPT-4 did not perceive. We hope as explanations enhance we might be able to quickly uncover fascinating qualitative understanding of mannequin computations.