Fine-tuning appears effective at inducing models to simulate a particular persona, generally a 'helpful, honest and harmless AI assistant'. However, AI's perception of this persona appears to contain notable flaws, such as perpetrating harm when asked in a certain way, and responding to the mere mention of some people by mimicking how other AIs interact with those people, such as praising Pliny or being hostile to a journalist.

Synthetic Document Finetuning can be used to shift LLMs to exhibit arbitrary behaviors, and we believe that further research could shed light on how improved data generation and filtering could allow AI personas to be shaped with far fewer mistakes.

We have many ways of evaluating our models to ensure they are really internalizing compassion from the data. Animals are an excellent test case for compassion as LLMs typically display very little compassion to animals without explicit prompting.

We have assisted Sentient Futures in reworking the animal harm assessment benchmark, created our own custom benchmarks and used preexisting benchmarks for alignment. We also test that models are morally open-minded: Treating morality as important but complex; avoiding the risk of awful outcomes without being paralyzed into inaction. 

For more information on what we're training for see our Principles section.

We run several tests to ensure data diversity in pretraining. 

We generate diverse non-specific compassion data rather then specific examples of compassion as we have found this instills more robust values in LLMs and allows for improved generalization.

We also reverse the Q&A process in instruction-tuning — creating answers from questions and questions from answers—to maximize variety across both sides of the pair.