English | 简体中文

Train your own Jev. Deploy it in your own application.

JevAny is open infra for training and serving Jev-style decision models. Fine-tune an open Qwen backbone on your labelled decisions, or start with our pretrained checkpoint. Both paths use the same Python and HTTP interface, following the Jev API's request and answer format.

Give the model a state, a question, and possible answers. It returns a choice and the probability of each option. Your application can route a message, select a query repair, or choose an agent's next action.

Start here	What JevAny provides
Train a model	Starter data, public-source data builders, SFT/RLCR recipes, and a multi-GPU training launcher
Deploy a checkpoint	Released checkpoints, local Python inference, and an HTTP server

See it act

Examples built with JevAny-27B-SFT:

Explore all 30 cases, with task descriptions and decision records. These are selected successful runs, not a success-rate benchmark. Three scenarios also have small runnable applications using the public interface.

Get the code

Python 3.12 or newer is required. Install the extras for the path you choose below.

git clone https://github.com/weitianxin/JevAny.git
cd JevAny
python3.12 -m venv .venv
source .venv/bin/activate

Deploy a checkpoint

The released 27B checkpoint needs a GPU that holds the full BF16 base model plus runtime memory; the showcase used A100 80 GB GPUs. First loading downloads the adapter and its separately distributed base weights. Smaller checkpoints you train use the same interface.

python -m pip install -e '.[serve,multimodal]'

jevany serve --checkpoint tianxinwei/JevAny-27B-SFT \
  --device cuda --dtype bf16 --port 8008

With the server running, make a decision from Python:

from jevany import Choice, JevClient, Noul

jev = JevClient("http://127.0.0.1:8008")
result = jev.system_one(
    state={"ticket": "I was charged twice. Please help."},
    questions={
        "department": Choice(
            instructions="Which team should handle this?",
            criteria={"billing": "Payment problems", "shipping": "Delivery problems"},
        ),
        "urgent": Noul(instructions="Does this require urgent review?"),
    },
)
print(result["answers"]["department"]["choice"])
print(result["answers"]["department"]["probabilities"])

For inference inside your application, load once with JevModel.from_pretrained(...) and call the same system_one method:

from jevany import JevModel

jev = JevModel.from_pretrained("runs/my-jev", model_name="my-jev")

You can also send JSON to POST /v1/systemone, run jevany decide examples/request.json, or point the official TypeSafe SDK at the server. The deployment guide covers each option, hardware, offline loading and media inputs. A client-only installation (pip install -e .) does not install PyTorch.

Train a model

Start with the included data and a small Qwen backbone:

python -m pip install -e '.[train]'

jevany data init --out data/starter
jevany data validate data/starter/train.jsonl
jevany train --config recipes/sft.toml --dry-run
jevany train --config recipes/sft.toml

# Serve the checkpoint you just trained.
python -m pip install -e '.[serve]'
jevany serve --checkpoint runs/my-jev --model-name my-jev

The starter contains 24 original synthetic training records and 8 separate development records. Each asks choice, binary and score questions. It is for learning the workflow; train on representative domain data to build a useful model.

The SFT recipe uses Qwen/Qwen2.5-0.5B and a CUDA GPU. Change base and data in the TOML file, or override them on the command line. Labels use the inference format with a label added to each question. The trainer fits a LoRA adapter and pointer head; the base weights stay frozen.

Next step	Command or guide
Train on your data	jevany train --config recipes/sft.toml --data data/my-domain.jsonl --out runs/domain-jev
Fine-tune our released checkpoint	recipes/finetune.toml
Build larger datasets	jevany data build-sft --help · sources and formats
Run on multiple GPUs or hosts	infra/train.sh
Experiment with calibration rewards	recipes/rlcr.toml

The training guide covers CPU overrides, Python training, evaluation, checkpoint selection and distributed launch. Training currently uses PyTorch, Transformers and PEFT with supported Qwen backbones. DDP keeps a full model on each GPU.

Run the examples

With a server running:

python -m examples.inbox
python -m examples.sql_repair
python -m examples.service_recovery

Example	What you can build from it
Inbox triage	Classify messages and decide which need a reply
SQL repair	Select a query, execute it in SQLite, and check its result
Service recovery	Run a multi-step agent against a local replica simulator

Pass --checkpoint runs/my-jev to run any example in-process. The SQL and recovery examples report failed checks as failures. Example instructions explain the environments; harness and symbolic integrations add optional LLM planning.

Checkpoints and evidence

Checkpoint	Role	Transfer accuracy
JevAny-27B-SFT	Default released model	82.41%
JevAny-27B-RLCR	Experimental RLCR continuation	82.31%

These v0.2 measurements cover 1,046 transfer questions. RLCR has not shown an overall transfer gain. Evaluation details include the full comparison, image/video controls and negative test-time adaptation results; ALGORITHM.md describes the training objective.

Native images and video require a compatible vision checkpoint and one question per request. HTTP media inputs are enabled explicitly through JEVANY_MEDIA_ROOT. The validated training window is 2,048 packed tokens. Calibration can change on new data; evaluate thresholds for your application.

Documentation and contributing

Training · Deployment · API compatibility · Data · Contributing · Research roadmap

JevAny is independent of Jev and TypeSafe and includes no Jev weights or private implementation. It includes infrastructure adapted from Kev; see NOTICE and ACKNOWLEDGEMENTS.md. Code and starter data are Apache-2.0. Base models and upstream datasets retain their own terms.