JevGate-9B

JevGate-9B is the admission gate of HealthJev, a self-evolving health agent with an external long-term memory. Before a distilled post-episode experience is written into memory, JevGate performs a fast, governed admission decision: only experiences that pass the gate are replayed and consolidated into the long-term memory layers (global rules, domain knowledge index, profile, and reusable task SOPs). It is designed as a lightweight System-One decision model, with a median admission latency of 86 ms as reported in the paper.

Architecture

• Backbone: frozen Qwen/Qwen3.5-9B-Base
• Adapter: LoRA (r = 16, alpha = 32, dropout = 0.05), initialized from jaredpalmer/kev-9b and trained for the admission objective
• Pointer head: 256-dim head predicting the admission score (head.pt)
• Loading: base model + released LoRA + released pointer head; no second upstream adapter

Files

File	Size	SHA-256
adapter_model.safetensors	173,178,096 B	cf1827f87615b821e78ae1f5b2957df79494829cda9209993d587186e55fc0a3
head.pt	8,393,087 B	e3f0b2b40be1f188e4d5baf59a852eaa3c943e7f350c413199ca9801f8e4d49f
adapter_config.json	1,309 B	613f8e0634a7f7e19d391067b66c0fb33cf9a4bc47aefbd2af30087b16d08bfe

release.json additionally records the per-task deployment thresholds and all checksums.

Usage

import torch
from huggingface_hub import hf_hub_download
from peft import PeftModel
from transformers import AutoModelForCausalLM

repo = "lihaoran51/JevGate-9B"
base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3.5-9B-Base", torch_dtype=torch.bfloat16)
model = PeftModel.from_pretrained(base, repo)
head = torch.load(hf_hub_download(repo, "head.pt"), map_location="cpu")

Deployment thresholds

The evaluation threshold is 0.5; per-task deployment thresholds used in the paper are:

Task	Threshold
diabetes_readmission	0.88
geneturing	0.38
odir5k	0.90
icu_sofa	0.50
mimic_icd	0.52
medcalc	0.30

Intended use and limitations

JevGate-9B is released for research on governed experience admission in self-evolving agents and for reproducing the HealthJev evaluation. It is a research artifact for decision support, not a medical device, and must not be used to make clinical decisions.

License

The adapter, pointer head, and configuration files are released under Apache-2.0 (see LICENSE). The base model and the upstream initialization checkpoint follow their respective licenses.

Citation

Code, benchmark, and evaluation suite: https://github.com/HC-Guo/Self-Evolving_HealthJev