Any2jev

It is a converter: take any Hugging Face causal LM plus a few thousand labelled decisions, train it, and get a model that answers Choice, Score and Noul questions in one forward pass with zero generated tokens. The README frames it as putting a Jev-shaped interface in front of another model, since Jev's own weights are closed.

The README's worked example finishes in about one GPU-hour: `any2jev train --base Qwen/Qwen3-0.6B --data train.jsonl --val val.jsonl --out runs/my-jev`. You need labelled decisions in JSONL plus a validation file for checkpoint selection.

Install with `pip install "any2jev[serve]"`, run `any2jev serve runs/my-jev`, and it exposes POST /v1/systemone. The README says the official TypeSafe SDK connects after a single base_url change; the project is Apache-2.0 and requires Python 3.10 or newer.

On a three-question support ticket with Qwen3-0.6B on an RTX 2060 SUPER: 42 ms per request, 0.796 accuracy and ECE 0.027, versus 778 ms and 0.621 for the same weights prompted for JSON. Accuracy and ECE come from 1,000 held-out questions (boolq, ag_news, banking77, sst5); the cloud row is deliberately blank because no cloud model was measured in the repo.

No key: it trains and serves your own local model. TypeSafe's closed Jev is the inspiration and the API shape, not a dependency. The README also notes that every number in its table comes from a script committed to the repository.

You supply both the base model and the labelled decisions, and training needs a GPU. The README estimates roughly 12 GPU-hours per 1M requests for the 0.6B example, and all numbers are fixture measurements from one machine, so treat the speedups as same-weights engineering comparisons rather than general benchmarks.

Choice, Score and Noul questions. The README lists routing, triage, moderation, ranking, guardrails and game agents as cases where you need a model that decides in milliseconds with probabilities your code can threshold, rather than one that writes.