Open-Jev

Checkpoints	Inputs
ZefanCai/Open-Jev-2B, -9B	text

An Open-Jev release is a LoRA adapter, a scalar head and a temperature on a pinned Qwen3.5 base. Flide asks it with the release's own prompts and reads each prompt at its last token with the scalar head. A choice or a score has one prompt per candidate and takes a softmax over their scores at the release's temperature; a noul has one prompt, and its probability is a softmax over zero and that score. The prompts of a request share their tokens up to where they first differ, and Flide computes that shared part once.

Prepare a checkpoint

Merge the adapter into its base once. The command downloads the release and the base it names, merges the adapter in float32, and writes a directory Flide loads directly.

python -m flide_open_jev.prepare ZefanCai/Open-Jev-2B open-jev-2b

Start a server

flide serve --model open-jev-2b

Send a request

curl http://localhost:8000/v1/systemone -H "Content-Type: application/json" -d '{
  "state": "Order #1182 arrived with a cracked screen. The customer asks for a refund and attached two photos.",
  "questions": {
    "action": {"type": "choice", "instructions": "What should support do next?",
               "criteria": {"refund": "refund the order", "replace": "send a replacement", "escalate": "ask a human"}},
    "urgent": {"type": "noul", "instructions": "Does this need a reply within the hour?"}
  }
}'

A choice takes up to 255 options and a score up to 10 levels. A candidate prompt is at most 4096 tokens unless the release sets another length. --format-arg max_length=8192 changes it.