Nokia open-sources AnyJev library to turn open LLMs into calibrated decision models
Nokia’s applied research team released AnyJev, an open‑source Python library that converts any open‑source large language model into a typed decision model without any training. The library reads the model’s next‑token distribution, applies cyclic‑shift averaging (L0) and batch prior correction to remove position and label bias, and optionally adds temperature scaling (L1) for larger label sets.
Key points
- AnyJev turns open LLMs into calibrated decision models without training, using cyclic shifts and batch prior correction.
- On Qwen3‑8B BANKING77, order‑flip rate fell from 0.230 to 0.073 and ECE dropped to 0.036 with L1.
- Library installs via pip, runs on Hugging Face or vLLM backends, and costs ~0.25 s per decision on an H100 (K=20, batch 32).
Benchmarks on Qwen3‑8B with the 20‑way BANKING77 test set show the order‑flip rate dropping from 0.230 to 0.073, and with L1 the expected calibration error (ECE) falls to 0.036 compared with 0.144 for the original Jev system. On a 100‑label task, auto‑decidable traffic at 5 % error rises from 7.7 % to 52.0 %. The library installs from PyPI, is Apache‑2.0 licensed, and supports Hugging Face and vLLM backends; a single decision with K = 20 costs about 0.25 s on an H100 GPU at batch size 32.
AnyJev aims at production scenarios where a single answer must be chosen from a fixed set, offering a fast, training‑free alternative to fine‑tuned classifiers.
The story so far
3 episodes →- Nokia open-sources AnyJev library to turn open LLMs into calibrated decision modelsthis story
Nokia Open-Sources AnyJev: A Training-Free Layer That Turns Any Open LLM Into a Calibrated Decision Model
MarkTechPost · 23 September 2026
Loading the full article…
This text was published by MarkTechPost and written by Asif Razzaq. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Agents & Tools
All →- Amazon blocks Meta's Muse AI agent from shopping on its platform · 11 src
- AI agents become malware distribution channel through fake GitHub repositories · 1 src
- Opinion: Claude adds Docs and Slides, sparking debate over office integration · 1 src
- Delinea tests Anthropic Claude Mythos 5.1 on its own privileged‑access code · 1 src
- Google launches CC family AI agent for up to six members · 1 src
Comments
via GitHub Discussions