Introducing Distribution Fine-Tuning
Deft is our first step toward productizing Distribution Fine-Tuning: a model-training approach for making generated writing match the shape of strong human prose.
We are beginning to productize a technical idea we call Distribution Fine-Tuning. Deft is the first product expression of that work: a writing editor that turns rough material into clearer, more human-feeling prose without asking the user to become a prompt engineer.
The problem starts with a familiar failure mode. Standard fine-tuning can make a model follow examples, but the outputs still drift away from the distribution of the training data. They overuse certain phrases, settle into repetitive sentence shapes, or trade blandness for incoherence when sampling settings are pushed higher.
Our technical report measures that failure directly. Token-distribution distance tracks overused words and phrases. Maximum Mean Discrepancy compares embedding distributions, which helps reveal whether model outputs are too generic or conceptually thin. Judge Model Quality asks a separate model to compare human and model outputs for writing quality.
Those metrics point to the same intuition: good writing is not only a sequence of locally plausible tokens. It is a distributional object. A finished essay has a texture, rhythm, density, and level of detail that can be missed when training only optimizes individual examples.
Distribution Fine-Tuning trains at that higher level. In the report, a DFT-trained 14B model improved MMD from 0.037 to 0.018 and Judge Model Quality from 0.49 to 0.80 against a strong supervised fine-tuning baseline. The same evaluation also showed large gains in creativity, coherence, clarity, and depth.
For Deft, the product implication is simple. The user should be able to bring a draft, outline, fragment, or generic model output into a calm editor, generate a stronger version, and keep revising. The model work is server-side. The visible product is the editing surface.
This is not a promise that a model can replace taste, research, or care. The early system is focused on web and essay-like prose, and the writer still owns the argument. But it changes the starting point: generated text can be trained to resemble the distribution of finished human writing instead of merely avoiding a few obvious bad habits.
That is the direction for Deft. Distribution Fine-Tuning is the model layer; the editor is the product layer. Together, they make writing with a model feel less like managing a chatbot and more like working inside a tool built for drafts.