technovangelist / notes / eta

eta

Eta is the learning rate in the context of Mirostat. Learning rate controls how rapidly Mirostat adapts to changes in the models output entropy. A higher rate means faster adaptation, lower is slower. the default value is 0.1. So if the next word chosen by the model is a strange choice, with a higher eta value, the model will continue to provide words that seem appropriate for this newer style. A low eta is more stable and less likely to overreact to temporary changes.

#ollama/parameters