Zero training. Ready to adapt.
Keep your model weights. Describe tasks with Choice, Score, and Noul to get structured decisions and probabilities.
NO FINE-TUNINGSkip to content01 / 02 / 03 / All project news 01 02 03 Get started
v0.1.0 · Ready to explore
YOUR MODEL. NEW POSSIBILITIES.
Give yourLLM / MLLMa decision engine.
Jev-style decisions. Zero training. Classify, score, and choose actions from text and images—with prefill alone.
L2J / DECISION FLOWPREFILL ONLY
Evidence
“My card was charged twice.”
Which department should handle this?Your existing model
0TRAININGPrefill → yes / no logits
Structured decision
Billing
0.92
Shipping
0.08
Illustrative probabilities · not live inference
RUN WITH THE BACKEND YOU KNOW
SGLang+Transformers+MLX
TEXT + VISIONLESS SETUP. MORE DECISIONS.
Put your existing model to work.
Decisions at prefill.
Read yes/no logits directly, without decoding an answer. Shared prefixes reduce repeated computation across candidates.
PREFILL ONLYChat & decisions. Same client.
Use the familiar OpenAI SDK interface. With SGLang, chat and decisions also share a service, inference engine, and loaded model.
OPENAI SDK EXTENSIONFAMILIAR INTERFACE. NEW CAPABILITIES.
A few lines. A clear decision.
Create a client, ask a question, get an answer with probabilities. Sync, async, text, and images are all supported.
Read the usage guideInstall your preferred backend first
uv sync --extra sglangdecision.py
from llm2jev import Client, Choice
client = Client(
base_url="http://localhost:30000/v1",
api_key="unused",
)
response = client.systemone.create(
model="local-model",
state="My card was charged twice.",
questions={
"decision": Choice(criteria={
"billing": "Charges and billing",
"shipping": "Delivery",
}),
},
)
print(response.answers["decision"])↳ OUTPUT · Illustrative probabilities · not live inference
{
"type": "choice",
"choice": "billing",
"probabilities": {
"billing": 0.92,
"shipping": 0.08
}
}FROM INPUT TO IMPACT
See. Decide. Act.
One decision interface. Different worlds to act in.
BUILDING IN THE OPEN
Latest updates
v0.1.0: give your LLM / MLLM Jev-style decision capabilities with zero training. Supports SGLang, Transformers, and MLX, with a Python client that extends the OpenAI SDK.
Valen Sokoban evaluation: added a visual Sokoban demo and single-step action evaluation based on Valen.
JevBench evaluation: LLM2Jev accuracy and latency on 231 public items; P50 latency is below one tenth of the Jev official result.
LLM2JEV / OPEN SOURCE
Your model. Ready to decide.
Open source · Apache 2.0 · SGLang / Transformers / MLX