Choose theme
AI Resources
AnyJev
AnyJev is a Python framework that uses an existing language model to return probabilities over supplied choices, yes-or-no answers, or score levels instead of generating prose.
Its raw, L0, L1, and L2 levels progress from direct token scores to zero-label bias correction, temperature calibration, and a small closed-form decision head fitted to one question and model. Use this as a first read, not a recommendation. Open the original project before trusting details like terms, limits, privacy, cost, setup, or safety.
What it is
Typed decisions from a model you already run
A builder supplies a question and bounded options. AnyJev reads next-token probabilities or an intermediate hidden state and returns an explicit distribution for choice, yes-or-no, or score decisions.
Why it stands out
A staged path for reducing decision bias
L0 uses no labelled examples and is designed to reduce option-order and label-prior effects. L1 adds temperature calibration, while L2 fits a small question-specific head without changing the underlying model weights.
Availability
Python package with two model backends
The Apache-2.0 project supports Transformers and vLLM, provides a pipeline command, tests, benchmark scripts with committed results, and a fake backend for trying the interface without downloading a model.
Why it matters
What makes it useful
A language model can give different choice probabilities when option order or label letters change. AnyJev makes those effects visible and offers explicit levels for correcting or calibrating the resulting distribution.
What to know
Where it fits
It fits local or self-hosted workflows that already have an open model and need bounded classification, routing, filtering, or scoring. Builders can begin with the zero-label L0 path, then test whether labelled calibration or a fitted head improves a repeated decision.
Notable points
What stands out
AnyJev is a new pre-alpha project. Its accuracy and backend-agreement figures are project-reported, and its typed-decisions accuracy measures agreement with a teacher language model rather than independent ground truth. The project also notes that each decision was tested in isolation rather than inside a longer agent loop.
Before using
What to review
Whether the task has a meaningful supplied option set. AnyJev scores bounded decisions and does not generate a missing answer.
Whether L0 helps on the real label distribution. It requires one prefill per choice and the project notes that its batch-prior correction can hurt on imbalanced labels.
Whether enough task-specific labels are available. The project suggests roughly 100 to 500 examples per question for L1 and 100 to 300 for L2.
Whether L2 fits the chosen model and serving path. Its head is specific to one question and model, requires an intermediate hidden state, and published heads currently cover five Qwen3 models.
Whether the present interface limits fit the task. The documented letter readout supports at most 26 options, and SGLang support is not yet available.
Whether project-run benchmarks transfer to the real workflow. Calibration cannot make an underlying model understand a task it cannot already perform.
Reader fit
Who may find it relevant
Builders testing a typed decision layer on top of an existing Transformers or vLLM deployment.
Teams that can evaluate option-order effects and, where useful, collect labels for a repeated decision.
Less relevant for readers who want a general chatbot, free-form generation, a managed service, or a ready-made L2 head for an unsupported model or question.
Editorial note
Why LifeHubber lists it
LifeHubber lists AnyJev because it makes a useful engineering choice inspectable: whether zero-label bias correction is enough for a bounded decision, or whether that decision justifies question-specific calibration or a fitted head on the model a builder already runs.
Source links
Source materials
Reader note
Before relying on this entry
LifeHubber lists entries to help readers inspect AI projects, not to endorse them or prove they are safe, suitable, accurate, maintained, or right for a specific use. We do not verify every entry in depth. Before relying on anything listed, review the original materials, terms, privacy practices, limits, and risks that matter for your situation.
More in Ecosystem
Keep browsing this category
Explore more AI ecosystem resources.
Laya
NandhaKishorM/laya
An early Apache-2.0 model family and Python runtime for bounded choice, score, and yes-or-no decisions, with English, multilingual, and task-specialized checkpoints plus a router that selects between them.
CLM
Contrastive-LM/CLM
An Apache-2.0 contrastive model and local server for typed decisions and candidate ranking, with separate state and action embeddings, reusable candidate caches, a browser playground, public reference heads, and fine-tuning tools.
FreeToken
FlashML-org/FreeToken
A local serving engine for running large mixture-of-experts language models across NVIDIA GPU memory, system memory, and CPU compute, with a local API, terminal interface, and coding-agent launch paths.
For project maintainers
Listed here? You can use the badge.
If you maintain a project with a current LifeHubber listing, you may add the optional “Listed on LifeHubber AI Resources” badge to its README, docs, or website. No introduction or permission request is needed.
