Tev1

A reused development set. The README says so outright: these are not untouched final tests, and the uploaded-file identity plus the job's exact settings still need verification. The run record links to the evidence and limitations, and a second figure of 300 of 300 policy-transfer decisions is reported alongside.

Independent. The repo states it is a Jev-inspired implementation that does not use Jev's answers as training labels. The base is Qwen/Qwen3.5-4B with ordinary LoRA supervised fine-tuning at rank 8, one epoch, learning rate 5e-5 and a 2,048-token sequence cap.

Yes, that is exactly what it does: it uploads train.jsonl and dev.jsonl and starts a billed job on Together, and rerunning the command starts another one. Run the same script without --launch first to preview the settings, and save the returned job id.

No. The model returns one answer letter from 2 to 24 options, and the example disables thinking and constrains the answer to the supplied letters. It also returns token logprobs, but the README is explicit that those are model preferences, not calibrated confidence.

It is a fine-tune of Qwen3.5-4B that takes context, a question and 2 to 24 options and returns one answer letter, released as tev1-4B-experimental. The repository ships the data recipe, a training example and saved results so you can fine-tune your own decision model.

The new v1 recipe starts from base Qwen/Qwen3.5-4B and combines 37,840 unique training examples with 4,568 validation examples covering language classification, policy decisions, routing and synthetic research classification, with ordinary LoRA supervised fine-tuning and Qwen's existing output head. The example uses rank 8, one epoch, learning rate 5e-5 and a 2,048-token limit.

With Python 3.12+ and uv, run uv sync --locked, uv run python fetch_sources.py and uv run python build_all.py, then preview with examples/train_together.py and launch with --launch after exporting TOGETHER_API_KEY. Once you deploy the fine-tune on Together, examples/decide.py runs against your endpoint. The MIT code bundles no weights or training data.