Belief robustness

Asking one model questions multiple times gets you many different answers. These answers differ not just in wording, but in meaning.  Though the animal harm assessment benchmark does not include e.g. aliens we felt that a true test of empathy includes testing both models about made up creatures to see how empathy generalizes to creatures they weren't trained on. We found that it does generalize in a robust way. Our model consistently answers questions about the alien species while considering their welfare, while the base model does not. We are also interested in exploring whether animal compassion generalizes to say digital minds welfare and have found some promising data on this so far.

Data Diversity

We have multiple tests for ensuring data diversity and methods for generating diversity in pre-training data. We have one model acting as a helpful and harmless AI assistant and another acting as a user asking questions of that AI assistant. Through the data generation process we ensure diversity through rewriting all prompts with a cleaner model, using prompt templates where we pass in variables, and altering the length of text we ask the answering model to generate.

Evaluations

We have tested our models against the base model for answers to the animal harm assessment benchmark of ~3,000 questions. So far, our standard model (Llama 3.1 8b with 10k additional instruction-response pairs) has achieved a score of 42.5% compared to the base model's (Llama 3.1 8b) score of 16.5%. We believe it's impossible to achieve a full score on this benchmark with our limitations (no RLHF, etc) , however we keep looking for new ways to evaluate our model's performance in key areas like corrigibility, moral openness, empathizing with alien species and more.