Jeva
NUS research System One: native encoder plus decision stack, LIQIIIII/Jeva-4B, Choice/Noul/Score with null awareness and option caching.
Overview
Jeva (github.com/LiQiiiii/Jeva, LIQIIIII/Jeva-4B on Hugging Face, arXiv:2610.18773) from Qi Li and Xinchao Wang at the National University of Singapore is a native System One stack. A language encoder represents state and each option independently; a compact decision head updates state against a complete option memory so probabilities stay permutation invariant, null aware, and scalable to variable candidate sets with two-level option caching. Python API exposes Choice, Noul, and Score via model.system_one or model.predict for dataset-exact option text. Project page liqiiiii.github.io/Jeva/ hosts figures; docs/evaluation.md and docs/finetuning.md cover benchmarks and training. jeva-doctor --check-kernels --strict validates the CUDA kernel install after pip install -r requirements.txt.
Problem: Teams want System One shaped Choice, Noul, and Score from a model designed for set-dependent decisions, not a chat model forced to emit one token per label.
Built for: Researchers and ML engineers with Linux x86_64 and NVIDIA GPUs who will run Jeva-4B locally, read arXiv 2610.18773, and fine-tune or evaluate using the repo docs.
First indexed on Jev Directory: 2026-10-15
Creator and team
- Name
- Qi Li
- Organization
- National University of Singapore
- Handle
- @LiQiiiii
How Jev is used
- Role in the product flow
- Native parallel Choice, Noul, and Score over encoded state and independently encoded options with reusable option tuples
- Primitives
- ChoiceScoreNoul
- State in
- Text state string plus questions map of Choice criteria dicts or lists, Noul instructions, or Score rubric criteria with optional values per README local_inference example.
- Decision out
- Per-question choice key and probabilities, noul truth probability, or expected score level; open-set Choice returns null_probability and conditional_probabilities when allow_null is true.
- conda env with CUDA 12.8 per README Installation
- Jeva.from_pretrained LIQIIIII/Jeva-4B on cuda
- Call system_one with parallel questions on one state blob
- Optional cache complete option tuples or reuse projected keys across new sets
- Use predict(...) when benchmark rows require exact option strings and order
Jeva is research weights and a training story, not a hosted api.typesafe.ai substitute. mapika-decider and yinsongxu-llm2jev ship wire-compatible HTTP servers from other Qwen lineages; Jeva publishes the encoder plus decision stack paper and HF Jeva-4B with finetune docs. tianyucodings-nanojev targets game-scale 0.6B heads; Jeva argues for first-principles structural properties on a 4B class model. nokia-applied-research-anyjev adds calibration pipelines on vLLM; featherless-simple-jev assembles classifier logits from arbitrary HF models. This directory does not claim TypeSafe affiliation for Jeva. Cross-read open model listings on /explore rather than duplicating NanoJev or AnyJev pages here.
Sourced performance claims
- README architecture tagline lists null awareness, permutation invariance, option scalability, and cacheable option memory as explicit design goals.Source: github.com/LiQiiiii/Jeva README
- Citation block lists arXiv preprint arXiv:2610.18773, Jeva: System One Models from First Principles, Li and Wang, 2026.Source: github.com/LiQiiiii/Jeva README Citation
- Public GitHub repo LiQiiiii/Jeva had 76 stars and 3 forks as of October 15, 2026; created 2026-10-12.Source: GitHub API, 2026-10-15
Features and stack
Features
- LIQIIIII/Jeva-4B weights with Choice, Noul, and Score in one system_one call
- Permutation invariant scoring and explicit null_probability on open sets
- Option tuple cache and per-option vector reuse documented in README
- docs/evaluation.md, docs/finetuning.md, and examples/local_inference.py
- PyTorch 2.9.1 and CUDA 12.8 paired requirements with jeva-doctor checks
Stack
- Python 3.12
- PyTorch 2.9.1
- CUDA 12.8 NVIDIA stack
- Hugging Face Hub
Pricing: Open research code and public weights; you pay for GPUs, electricity, and Hugging Face bandwidth. No TypeSafe API required for local inference.
Links
FAQ
- Is Jeva affiliated with TypeSafe?
- No. Jeva is an independent NUS research release with its own GitHub repo, Hugging Face weights, and arXiv paper. This listing does not imply TypeSafe endorsement.
- Which Hugging Face model id should I use?
- README Installation and examples use LIQIIIII/Jeva-4B with Jeva.from_pretrained on cuda.
- How does Jeva differ from LLM2Jev?
- LLM2Jev serves POST /v1/systemone from general causal or VLM logits on SGLang, Transformers, or MLX. Jeva trains a dedicated encoder and decision stack with published structural properties and finetune docs.
- What hardware does the README require?
- README Installation: Linux x86_64, NVIDIA GPU, conda environment with cuda-nvcc 12.8.13 and pip requirements.txt; first request may compile kernels.
- Where is the paper?
- README Citation lists arXiv:2610.18773, Jeva: System One Models from First Principles, by Qi Li and Xinchao Wang, 2026.
Related learn guides
Original Jev guidance that pairs with this product pattern.
- System One model
The model family behind Jev: parallel typed questions, one forward pass, probabilities you can threshold.
- Jev vs LLM classification
When to gate with System One probabilities instead of asking a chat model to label things.
Related products
Hand-picked neighbors with rich profiles or overlapping tags.