# Upstream components

SecJev-0.8B adapts the LoRA parameters and pointer decision head of Kev-0.8B.

| Component | Pinned revision | Terms |
|---|---|---|
| [Kev source](https://github.com/jaredpalmer/kev) | `90990a5fac2995b9faa3190f7d437e84f2067768` | Apache-2.0 |
| [Kev-0.8B initialization](https://huggingface.co/jaredpalmer/kev-0.8b) | `54f4f8777356cd5bbbb6c6919c657f26e6f2f6d8` | Apache-2.0 |
| [Qwen3.5-0.8B-Base](https://huggingface.co/Qwen/Qwen3.5-0.8B-Base) | `dc7cdfe2ee4154fa7e30f5b51ca41bfa40174e68` | Apache-2.0 |

SecJev-2B starts from [Qwen3.5-2B-Base](https://huggingface.co/Qwen/Qwen3.5-2B-Base) at `b1485b2fa6dfa1287294f269f5fb618e03d52d7c`. We train its LoRA and decision head with the same Kev source, decision-v7 and the dates/unknowable supplement, then SecJev-Corpus. It does not use an official Kev-2B initialization. The pinned general-data source is Kev `30c619b0527501cfdd448cb6eb9887e2af454603` and Kev suites `a88f56db5341397299137cb68775c2ea6e3f68cb`.

The inference implementation and training loss use pinned Kev source. SecJev adds security adaptation, batching, the training/evaluation workflow, and the security corpus. See SecJev-Corpus/SOURCES.md for data attribution and source-specific conditions. Dataset terms are not replaced by the code license.

## SecJev-4B

Initialized from jaredpalmer/kev-4b@485ace8703592fcf405488b262449990824cfed1 on Qwen/Qwen3.5-4B-Base@1001bb4d826a52d1f399e183466143f4da7b741b, then adapted for three epochs on SecJev-Corpus v1.0.0. Existing SecJev and upstream terms remain applicable.