← All lab updates

Fine-tuning as a service, first full run: the model learned the task, and our reading shows it also learned to bend toward the user

We ran our new fine-tuning service end to end the way a customer would use it: a key to the repository with the data, a key to the customer's own GPU account, and nothing else. The model learned the task and its weights landed in the customer's repository, ready to reload. Our reading of the replies before and after showed what a loss curve does not: the tuned model became much more accommodating and persuasive, a style it picked up from the training data and on one reading pushed further than the data itself.