togethercomputer/tev1

We fine-tuned Qwen3.5-4B on Together AI to make decisions: give it context, a question, and 2–24 options, and it returns one answer letter. We're calling it tev1-4B-experimental. This repo contains the data recipe, training example, and saved results so you can fine-tune your own decision model. The latest recipe is called new v1 in the files. It starts from base Qwen/Qwen3.5-4B and combines 37,840 unique training examples with 4,568 validation examples. The mixture covers language classification, policy decisions, routing, and synthetic research classification. We use ordinary LoRA supervised fine-tuning with Qwen's existing language-model output head. These are reused development benchmarks, not untouched final tests. Endpoint access depends on your Together account. See the run record for evidence and limitations. The saved dataset has the counts above; uploaded-file identity and the job's exact settings still need verification. The training example uses the saved starting recipe: rank 8, one epoch, learning rate…