diff --git a/README.md b/README.md index f7e52ebe..f3de82bb 100644 --- a/README.md +++ b/README.md @@ -192,6 +192,7 @@ The [independent model research](community/projects/tools/README.md#independent- - [agy-jevgate](https://github.com/catpotd/agy-jevgate) - Fail-closed Antigravity `PreToolUse` hook for shell commands using TypeSafe Jev risk scores. [Project guide](community/projects/tools/agy-jevgate.md). - [AlphaOptimizer](https://github.com/alpha-tales/alphaoptimizer) - Compact large Codex/tool outputs locally; optional TypeSafe Jev ranks relevant chunks (BYOK). [Project guide](community/projects/tools/alphaoptimizer.md). - [agent-desktop](https://github.com/lahfir/agent-desktop) - Rust macOS accessibility CLI for desktop computer use; optional jev-desktop skill/scripts use TypeSafe Jev for target/command choice without putting the a11y tree in agent context (BYOK). [Project guide](community/projects/tools/agent-desktop.md). +- [AnyJev (Nokia Applied Research)](https://github.com/nokia-applied-research/AnyJev) - Turn open LLMs into Jev-style typed decisions with probabilities (L0–L2); independent of official Jev / TypeSafe hosted API. [Project guide](community/projects/tools/nokia-anyjev.md). - [feelings](https://github.com/BoundaryML/feelings) - BAML `.feels()` / `.how()` / `.matches()` typed AI-if methods powered by TypeSafe Jev; upstream licensing unspecified. [Project guide](community/projects/tools/feelings.md). - [jev-feels](https://github.com/Qew7/jev-feels) - Ruby gem for TypeSafe Jev as `feels?` / `decide` / `score` and Rails validations (distinct from BAML feelings and ruby_decision_model). [Project guide](community/projects/tools/jev-feels.md). - [jev-foundation-models](https://github.com/peterfriese/jev-foundation-models) - Swift 6 bridge: TypeSafe Jev as an Apple Foundation Models `LanguageModel` for `@Generable` decisions (distinct from TypeSafe Swift client). [Project guide](community/projects/tools/jev-foundation-models.md). diff --git a/community/projects/tools/README.md b/community/projects/tools/README.md index 53dccf36..d22789ea 100644 --- a/community/projects/tools/README.md +++ b/community/projects/tools/README.md @@ -286,6 +286,7 @@ These projects study related typed-decision patterns using other models. They ar | Project | What you can do | Stack / format | | --- | --- | --- | +| [AnyJev (Nokia Applied Research)](nokia-anyjev.md) | Turn an open LLM into Jev-style typed decisions with probabilities (L0–L2); independent of official Jev. | Python · PyPI (`anyjev` 0.1.0, Apache-2.0) | | [Jev-Omni](jev-omni.md) | Run an open multimodal System One–style classifier (text/image/audio/video → option probabilities) on Gemma 4 12B IT; independent of official Jev. | Python / PyTorch · HF weights (CUDA, ~50 GB FP32) | | [Jevlike](jevlike.md) | Train a small option-attention scorer with synthetic data and optional frozen encoders; independent of official Jev. | Python / PyTorch · research starter | | [NanoJev](nanojev.md) | Study independent Qwen-based typed decision heads, local serving, and game controllers with recorded comparisons. | Python / PyTorch · model research and replay | diff --git a/community/projects/tools/nokia-anyjev.md b/community/projects/tools/nokia-anyjev.md new file mode 100644 index 00000000..a4534d71 --- /dev/null +++ b/community/projects/tools/nokia-anyjev.md @@ -0,0 +1,49 @@ +# AnyJev (Nokia Applied Research) + +[All projects](../README.md) · [Independent model research](README.md#independent-model-research) + +Turn an open LLM into a Jev-style typed decision model (Choice/Noul/Score with probabilities) via position-debiasing and optional calibration/heads—no TypeSafe hosted Jev and no fine-tune required for L0. Distinct from any MorrisZJ/AnyJev fork; independent of official Jev weights. + +| At a glance | Details | +| --- | --- | +| Source | [Source](https://github.com/nokia-applied-research/AnyJev) | +| Maintainer | [nokia-applied-research](https://github.com/nokia-applied-research) (authors: Jiamu Zhang, Tianze Yang, Yucheng Shi, Liang Wu). Independently curated; this entry is not an upstream submission or endorsement. Not affiliated with TypeSafe AI. | +| Format | Python package **`anyjev` 0.1.0** (PyPI; `pip install "anyjev[hf]"`). Transformers backend today; vLLM/SGLang on roadmap. | +| Requirements | Python ≥ 3.10; Hugging Face model weights for live LLM readout (`torch`/`transformers` extras). Offline `demo` with `--backend fake` needs no download. | +| License | [Apache-2.0](https://github.com/nokia-applied-research/AnyJev/blob/3cd8c6fcd9e90fc04214575ade6779da1e3f3704/LICENSE). | +| Disclosure | AI-assisted catalog review; no affiliation. Listing is not an endorsement. Source inspected (README, LICENSE, pyproject, package layout). Live model downloads and reported BANKING77 numbers were **not** reproduced on the review host. Upstream benchmarks are author-reported. | + +## When to use + +Use it to study or deploy Jev-shaped decisions on your own open models with L0 (training-free), L1 (temperature), or L2 (closed-form early-exit heads). Prefer hosted [TypeSafe Jev](https://docs.typesafe.ai) when you want the official System One API; prefer [Open Alternative to Jev](open-alternative-jev.md) for other open-model comparison labs. + +## How it works + +`Decider` + backend read next-token distributions for typed questions. L0 averages cyclic option rotations and divides out a label prior; L1 adds temperature; L2 fits a small head on hidden states (~⅔ depth) from a few hundred labels. Every `Decision` carries its `level` for downstream gating. + +## Get started + +```sh +pip install "anyjev[hf]" +# or from source: +git clone https://github.com/nokia-applied-research/AnyJev.git +cd AnyJev +git checkout 3cd8c6fcd9e90fc04214575ade6779da1e3f3704 +python -m demo.jev_mode --backend fake # no download +``` + +## Examples and demos + +- README usage with `Question.choice` / `noul` / `score`. +- `demo/` lifecycle scripts; shipped heads under `anyjev-heads/`. +- Bench docs: `docs/results_bench.md` (author-reported). + +## Limits and data handling + +Local inference sends prompts only to your chosen backend—not to TypeSafe. Not a drop-in replacement for official Jev behavior or pricing. Pre-Alpha classifiers. This listing did not download Qwen weights or re-run benches. + +## Review and maintenance + +Reviewed on **2026-09-23** at [commit 3cd8c6f](https://github.com/nokia-applied-research/AnyJev/tree/3cd8c6fcd9e90fc04214575ade6779da1e3f3704) (**0.1.0**, Apache-2.0). AI-assisted source review of README, LICENSE, pyproject. No live model spend. + +Related: [Open Alternative to Jev](open-alternative-jev.md), [Jev-Omni](jev-omni.md), [SemIf](semif.md).