GitHub · Research and data
nokia-applied-research/AnyJev
The cheapest way to test the decision-model thesis on a model you already serve.
What it does
Turns an off-the-shelf open LLM into a model that answers typed questions — choice, noul, score — by reading the probabilities of its first output tokens instead of letting it write. Ships pre-built calibration heads for five Qwen3 sizes, and reports that a 1.7B at 64% of its depth reaches the accuracy Jev publishes.
The repository's own description: “Turn any LLM into a Jev-style decision model: typed decisions, real probabilities, no training. (continue updating, welcome any issue and PR request)”
How it works
A small trained head sits on the model's logits and maps them onto the question the caller asked, so a single forward pass produces the label and its probability. Nothing is generated, and the evaluation scripts ship beside the heads so the table can be re-run.
The repository, by the numbers
Read from the GitHub API on 2026-09-27. Stars and the last push move daily — quote them with the date, the way we do.
What we checked
- the README's results table and the per-cell document it links
- the licence (Apache-2.0) and the pip package name
- the head sizes listed under anyjev-heads/
- who wrote it: Nokia Applied Research, with a Tencent Hunyuan co-author
What we did not check. We did not install it or run any of the reported evaluations. The order-flip and calibration figures are the authors' own on their dataset, and the comparison against Jev and Laya is against those projects' published numbers, not a head-to-head run.