How to train your own Jev for $17

Together AI has put out a guide for building a Jev-style classifier on top of Qwen3.5 4B, and the pitch is as blunt as the price tag: about $17 to train, then a deployed model you can query over HTTP.

The job of this kind of model is narrow and useful. Give it some state plus a question and a fixed set of answers, and it returns a score, a yes-or-no, or a choice. In the example, a customer says they were charged twice for an October subscription, and the model is supposed to pick the right support intent from four options.

To get there, Together uses 37,840 training examples drawn from six datasets, including MultiNLI, BoolQ, Banking77, AG News, SST-5, and a couple of others focused on rules and paper classification. The company says that dataset size keeps costs low, while larger runs get more expensive and take longer.

The workflow is pretty direct. Clone the tev1 repo, set up the environment, fetch and normalize the data with the provided scripts, then launch a fine-tuning job through Together’s service. The blog says the training job takes roughly 25 minutes, after which you retrieve the model name and create a dedicated endpoint on hardware labeled 1x_nvidia_h100_80gb_sxm.

And because the model was trained on JSON input and output, the calling pattern matters. Together says its example script hides some of the fiddly bits, but direct API calls need temperature set to 0, max_tokens to 8, and thinking turned off. Once deployed, the model can answer classification tasks like support intent, yes/no comprehension, policy checks, and sentiment analysis, then be shut down again when you’re done.