Project profile
AnyJev
Nokia Applied Research library that turns an existing LLM into typed decisions, with a no-label mode and a small fitted head.
Nokia Applied ResearchGitHubDocumentation
Last verified: 2026-09-30
Quick facts
| Project | AnyJev |
|---|---|
| Creator | Nokia Applied Research |
| Type | Library |
| Language | Python |
| License | Apache-2.0 |
| Uses Jev for | Training-light decision layer on someone else's LLM |
| Open source | Yes |
| Status | Active |
What it is
AnyJev is an Apache-2.0 Python package. The README shows a Decider over a vLLM server. Level L0 rotates options so position does not pick the winner. Level L2 fits a closed-form head on hidden states when you have labels.
Why it matters
It separates two different jobs that product copy often merges: debiasing a readout with no labels, and calibrating one after you label a few hundred examples.
Relationship to Jev
| Hosted Jev | Not called. |
|---|---|
| API | Its own Decider. The reviewed README serves vLLM, not the TypeSafe Python SDK. |
| Without labels | L0 with adaptive option rotations. The README says the stop rule is measured against its own full readout. |
| With labels | L2 fits a head. The quick start comment says on the order of 100 to 300 labels. |
Architecture
- Open LLM
- Hidden state or logits
- AnyJev head
- Distribution over your options
Key features
- pip install anyjev, with a Hugging Face extra.
- The README says an L2 deployment is a pooling server plus a few kilobytes of head.
- Authors listed on the README are at Nokia, Sunnyvale, and one author at Tencent Hunyuan.
Performance and claims
The README banner reports an order-flip rate moving from 0.230 to 0.073 with zero labels, and a calibration error moving from 0.240 to 0.095 with 100 to 500 labels.
Jev AI Hub read the README quick start and the repository record on 2026-09-30. It has not reproduced those rates.
Read AnyJev as a library with two operating points, L0 and L2, not as a single accuracy number against Jev. The no-label mode removes a position bias the authors measured. It does not by itself certify a production threshold.
Limitations
- You still run the underlying LLM. AnyJev does not replace that model's weights with a 400M encoder.
- The banner's before/after rates are the project's own measurements.
- This hub has not fit a head or replayed the order-flip demo.
Code and repository
https://github.com/nokia-applied-research/AnyJev
GitHub snapshot captured 2026-09-30: 974 stars, 126 forks, 12 open issues. Last push 2026-09-28. Counts change; this is not a live lookup.
Related Jev concepts
Related guides
Related projects
Sources
- AnyJev READMENokia Applied Research · accessed 2026-09-30 · github
- nokia-applied-research/AnyJev repository metadataGitHub · accessed 2026-09-30 · third-party