How to train your own Jev for $17
Short summary
Together AI walks through building and deploying a Jev-like classification model on Qwen 3.5 4B for $17. The guide covers dataset selection, fine-tuning, and deploying the model behind an HTTP endpoint so you can query it for tasks like intent detection or sentiment analysis.
- You can fine-tune a working classification model on 38,000 examples for roughly $17 and have it deployed in 25 minutes using Together's platform
- The model accepts JSON-formatted state plus predefined questions and returns structured answers (scores, booleans, or multiple choice), making it practical for e-commerce returns, paper categorization, or sentiment rating
- The entire process is automated through provided scripts—you clone a repo, download datasets, run a training command, and get back a deployed API endpoint ready to integrate
Full summary
Together AI just released their own Jev-like classifier built on Qwen3.5 4B, and they've published a complete walkthrough for fine-tuning your own version. Jev is a classification model that takes some context and predefined questions, then returns scores, boolean values, or multiple choice answers—useful for things like evaluating customer returns, categorizing papers, or rating sentiment. The tutorial walks you through the entire pipeline: cloning their GitHub repo, setting up dependencies and API keys, then moving into the actual training phase.
The training process uses six datasets totaling 38,340 examples sampled from Hugging Face, deliberately kept small to keep costs down—the whole fine-tuning runs about $17. You normalize the data using their Python scripts, then launch the training job through Together's service, which takes roughly 25 minutes. Once training finishes, you get a model name that you use to deploy a dedicated HTTP endpoint, which exposes your fine-tuned model so you can query it from any software.
The real test comes when you actually use it. They show an example where a customer complains about being charged twice, formatted as JSON with the question and possible answers. You send this to your deployed model using their provided script, and it correctly identifies the issue and returns the answer in the exact JSON format it learned during training. After that, you can either keep the endpoint running or shut it down and restart it later—or just use their pre-trained version on their serverless platform if you want to skip the training altogether.