JEV-mini
Minimal training project for a Jev-style decision model: Qwen3-0.6B with a pointer head that outputs probabilities over options.
Overview
jev-mini is a small learning project for training a decision model that outputs probabilities over candidate options. It uses a Qwen3-0.6B decoder with a classification-style PointerHead, and the README says it was extracted from Kev and is not an official Jev implementation. The author notes that a decoder plus head is redundant compared with encoder architectures like BERT and plans a simpler version and a multimodal one.
Training data is one JSON object per line with question types choice (option probabilities), noul (probability of true) and score (expected rating), and a distillation format with soft targets is supported. Configs control context limits, validation and evaluation; outputs include the adapter, pointer head and tokenizer but not base weights. The README shows a real tool-selection validation sample choosing among database_query, api_call, file_read and web_search.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which tool should the pipeline use? | database_query · api_call · file_read · web_search |
| Yes / no | Probability that the statement is true | yes · no |
| Score | Expected rating on a scale | — |
How it works
- Prepare local Qwen3-0.6B weights and JSONL training data
- Train with python -m jev_mini.train and the YAML config
- Evaluate with run_eval.sh on a test set
- Predict probabilities for a request with jev_mini.predict
Strengths
- Small, readable codebase for learning how decision heads work
- Supports soft-label distillation targets
- Saves only the adapter and head, keeping outputs small
Caveats
- Learning project; architecture acknowledged as redundant
- Training scripts contain machine-specific paths
- No licence stated
Getting started
python -m pip install -e . then python -m jev_mini.train --config config/train.yamlAt a glance
- Type
- Model
- Status
- research
- Language
- Python
- Runtime
- PyTorch training scripts
- Has code
- yes
Use cases and integrations
How curated lists describe it
JEV-mini is a minimal learning project to train a fast decision-making model.
wh000wh000/awesome-jev-live
Frequently asked questions
Is this an official Jev model?
No. The README says it is extracted from Kev and not an official Jev implementation.
Which question types does it support?
choice, noul and score, plus a distillation format with soft targets.
What base model does it use?
Qwen3-0.6B with an added pointer head.
Related projects
33-chapter Chinese tutorial on agentic RL and GRPO, including a chapter on when Jev-style external decision models beat training with RL.
ProjectresearchPythonInstruction dataset of 119 rows compiled from TypeSafe's public Jev documentation, split by Choice, Noul and Score.
ProjectresearchPythonKorean study notes and TypeScript mock-ups of Jev concepts, with a local Laya versus GPT-5 comparison and a calibration toolkit.
ProjectresearchTypeScriptFrom-first-principles PyTorch reconstruction of a Jev-like architecture with typed noul, choice and score heads; random weights only.
ProjectresearchPythonLearning project that builds choose-an-answer decision models from scratch and from Qwen, with downloadable 0.8B and 4B releases.
LibraryresearchPythonPreprint survey of Jev-style typed decision models: their research roots, adaptation costs and what current evidence supports.
Modelresearch