SecJev v1.3.0 — SecJev-9B

SecJev-9B is now available, trained from official Kev-9B on SecJev-Corpus for three epochs. Full development selected epoch 3; calibration used a separate split.

Downloads

• SecJev-9B-v1.3.0.tar.gz — trained LoRA weights, FP32 decision head, calibration, tokenizer, inference entry point and examples. The pinned Qwen3.5-9B-Base backbone is downloaded separately by bootstrap.
• SecJev-code-v1.3.0.tar.gz — training, inference and evaluation code for all released sizes.
• SHA256SUMS.txt — download checksums.
• SecJev-Corpus v1.0.0 — unchanged dataset.

Results

All 21,634 security test questions, 14 tasks:

Metric	SecJev-9B
Micro accuracy	88.11%
Macro-task accuracy	94.51%
Macro-task balanced accuracy	92.19%

Security performance is close to 4B. Kev clean-suite accuracy is 87.33% on decision-v7 (original 87.50%) and 84.76% on transfer-v4 (original 85.06%). Per-task results and all three checkpoints.

Usage

See inference instructions and nine-GPU training recipe.

9B uses a BF16 backbone with original FP32 LoRA/head, unmerged. Actual package inference and same-batch export parity were checked on an RTX 3090. The existing SecJev agreement and applicable upstream terms remain in effect.

SecJev v1.2.0 — SecJev-4B

SecJev-4B is now available: a security decision model adapted from official Kev-4B for three epochs on SecJev-Corpus v1.0.0.

Epoch 3 was selected on the full development split before testing. On all 21,634 security test questions:

Metric	SecJev-4B
Micro accuracy	88.01%
Macro-task accuracy	94.51%
Macro-task balanced accuracy	92.42%

Original Kev clean suites: 86.58% on decision-v7 and 83.84% on transfer-v4.

• SecJev-4B-v1.2.0.tar.gz: actual trained LoRA weights, decision head, tokenizer, calibration, inference code and model card. The pinned Qwen3.5-4B-Base backbone is downloaded separately using the included bootstrap.
• SecJev-code-v1.2.0.tar.gz: inference, training and evaluation source for the released models.
• SHA256SUMS.txt: file integrity checks.

Inference instructions · 14-task results and all three checkpoints · Nine-GPU training recipe

The unchanged SecJev-Corpus v1.0.0 contains 170,185 questions. The 0.8B and 2B releases remain available. Research-split scores are not live deployment detection rates. The existing SecJev agreement and applicable upstream terms apply.

SecJev v1.1.0 — SecJev-2B

SecJev-2B is now available: trained LoRA weights, decision head, calibrated probabilities, tokenizer and inference code. The pinned Qwen3.5-2B-Base backbone is downloaded separately by the included bootstrap script.

The model first learns general decisions from Kev’s method and data, then trains for three epochs on SecJev-Corpus v1.0.0. Epoch 3 was selected using the full development split; calibration uses its own split. This is our own 2B run, not an official Kev-2B checkpoint.

Security test metric	SecJev-2B
Micro accuracy (21,634 questions)	86.48%
Macro-average accuracy (14 tasks)	92.98%
Macro-average balanced accuracy	90.31%

Original Kev clean suites: decision-v7 84.50%; transfer-v4 75.91%.

• SecJev-2B-v1.1.0.tar.gz: actual trained model adapter/head, inference scripts, model card, selection and evaluation records.
• SecJev-code-v1.1.0.tar.gz: repository source, including 2B training and evaluation.
• SHA256SUMS.txt: download checksums.

Inference instructions · Per-task results and all checkpoints · Training recipe

The corpus is unchanged: SecJev-Corpus v1.0.0. The 0.8B release remains available. Existing SecJev and upstream terms continue to apply.

Package checks confirm unchanged adapter bytes, head tensors and temperature, plus CPU/GPU prediction agreement on six development cases across all three question types. This is packaging validation, not a new accuracy benchmark. Known limitations, including weak vehicle-message tasks and a failed handwritten policy example, are documented in the model card.

SecJev v1.0.0 — 0.8B model, corpus and code

SecJev-0.8B is a security decision model fine-tuned from Kev-0.8B. This release includes the trained model, the complete SecJev-Corpus and the training/inference code.

Downloads

• SecJev-0.8B-v1.0.0.tar.gz: trained LoRA weights, decision head, calibration parameters and tokenizer. The pinned Qwen3.5-0.8B-Base backbone is downloaded separately by the setup script.
• SecJev-Corpus-v1.0.0.tar.gz: all 170,185 questions in training, calibration, development and test splits, with source metadata and terms.
• SecJev-code-v1.0.0.tar.gz: inference, initial training, two-epoch continuation, development checkpoint selection, calibration, test evaluation and export.
• SHA256SUMS.txt: download checksums.

Results

SecJev-0.8B reaches 90.34% macro-average accuracy, 88.00% macro-average balanced accuracy, and 85.03% micro accuracy across 14 security tasks. The epoch-3 checkpoint was selected on development data. Full per-task results and original-Kev capability retention results are in the repository.

The model is released as an adapter plus decision head, not a full copy of the Qwen backbone. Usage and source-specific terms are in LICENSE, SOURCES.md and THIRD_PARTY_NOTICES.md.

See INFERENCE.md to run the model and TRAINING.md to reproduce training.