Standard LLM training is surprisingly bad at making the model outputs match the training data distribution, so the writing quality is bad.

I made a new training algorithm called Distribution Fine Tuning (DFT) to fix this.

The demo lets you try out a model trained with DFT.

More details in the technical report: https://rosmine.ai/2026/05/18/fixing-llm-writing-with-distri...