SecJev-4B

Mirror note (macmacmacmac): This repository is an unmodified mirror of the upstream SecJev-4B v1.2.0 release from UESTC1010/SecJev, redistributed under section 3 of the SecJev Use and Distribution Agreement v1.0 with LICENSE, notices, and provenance retained. The mirror exists so the jev-stageA JEV agent-defense evaluations can fetch both required checkpoints (this checkpoint and macmacmacmac/SecJev-4B-Action-Monitor-C1) from the Hub. Cite the upstream SecJev project for the checkpoint. The original upstream model card follows unchanged.

SecJev-4B scores Boolean judgments, choices and ordered ratings from security evidence and explicit policies. It returns probabilities over candidate answers instead of generating answer text.

This release contains the trained LoRA adapter, decision head, tokenizer, calibration and inference code. Download the pinned Qwen3.5-4B-Base backbone separately with the included bootstrap. We adapted official Kev-4B on SecJev-Corpus for three epochs; full development selected epoch 3. Temperature 1.3349681034 is fitted on calibration only.

Run

python -m pip install -r SecJev-4B/requirements.txt
export SECJEV_WORK="$PWD/secjev-4b-work"
python SecJev-4B/bootstrap.py --inference-only
python SecJev-4B/infer.py --model ./SecJev-4B --base "$SECJEV_WORK/base" --request SecJev-4B/examples/certificate.json --device cuda

The included example predicts reject; the actual answer is in example-output.json. Six short development cases (two per type) confirm export parity. They are a packaging check, not an accuracy estimate.

Data and results

SecJev-Corpus v1.0.0 has 170,185 questions from eight sources and 14 tasks, split into 118,818 training / 14,821 calibration / 14,912 development / 21,634 test questions. Download the unchanged corpus.

On all security test questions: 88.01% micro accuracy, 94.51% macro-task accuracy, 92.42% macro-task balanced accuracy. Original Kev clean suites: 86.58% decision-v7 and 83.84% transfer-v4. Full per-task results and all three checkpoints. Vehicle tasks remain weaker; research splits were inspected during corpus construction and are not a new blind deployment benchmark.

Training and terms

Nine RTX 3090 GPUs, three epochs, LoRA r16/alpha32 across attention/MLP/DeltaNet, 256-dimensional head, FP32 masters and BF16 autocast. Reproduction code. Exact initialization and selection records are included in provenance.json and security-training-config.json.

The SecJev Use and Distribution Agreement v1.0 and applicable upstream source/model terms apply. See LICENSE, SOURCES.md and THIRD_PARTY_NOTICES.md. This release does not replace the project agreement with Apache-2.0.