| --- |
| license: apache-2.0 |
| language: |
| - en |
| - zh |
| - es |
| library_name: kev |
| base_model: |
| - Qwen/Qwen3.5-9B-Base |
| - jaredpalmer/kev-9b |
| pipeline_tag: text-classification |
| tags: |
| - decision-model |
| - qwen3.5 |
| - kev |
| - choice |
| - calibration |
| - local-serving |
| --- |
| |
| # OpenJev 9B |
|
|
| OpenJev 9B is an open, locally deployable typed decision model for `choice`, |
| `noul` (yes/no/unknown-style decisions), and ordinal `score` questions over a |
| shared state. It uses Qwen3.5-9B-Base with Kev's pointer-style decision head and |
| a local Go API boundary that isolates sibling questions. |
|
|
| This release does **not** claim parity with hosted Jev. No Jev output was used |
| as a training label, soft target, synthetic teacher, calibration target, |
| selection signal, or fallback response. No fresh same-item Jev evaluation was |
| run for this checkpoint. |
|
|
| ## Identity |
|
|
| | Field | Value | |
| |---|---| |
| | Model | `openjev-9b-hard-final-w25` | |
| | Base | `Qwen/Qwen3.5-9B-Base@68c46c4b3498877f3ef123c856ecfde50c39f404` | |
| | Kev code | `jaredpalmer/kev@0fe8fc97c2bcc247fa3efb6e5c32af4e99770e91` | |
| | Weights SHA256 | `67b87afe12689bf6a95f6669d64250eae70e8744d0700019220942396355f8b8` | |
| | Head SHA256 | `56910f6a89bfb0cc5642764d7f983f5bad3d025dabcbad38590765455e7622ad` | |
| | Temperature | `1.9543199368684914` | |
| | Precision | BF16 | |
| | Maximum state | 16,384 tokens in the validated serving contract | |
| | Question semantics | strict isolated execution | |
|
|
| The four `model-*.safetensors` files contain merged full weights. `head.pt` |
| contains the typed Kev head. `SHA256SUMS` binds every checkpoint file. |
|
|
| ## Run locally |
|
|
| Use the exact Kev revision for the reference backend: |
|
|
| ```bash |
| git clone https://github.com/jaredpalmer/kev.git |
| cd kev |
| git checkout 0fe8fc97c2bcc247fa3efb6e5c32af4e99770e91 |
| uv sync --extra serve |
| KEV_TEMPERATURE=1.9543199368684914 \ |
| uv run --extra serve python -m kev.serve \ |
| --run iFurySt/openjev --port 8008 |
| ``` |
|
|
| The Kev endpoint implements the System One-shaped decision API. The OpenJev |
| source tree additionally contains a Go boundary with input validation, |
| authentication, overload handling, question isolation, readiness, benchmarks, |
| and an offline deployment layout. It has no Jev key or cloud fallback path. |
|
|
| ## Training method |
|
|
| The selected checkpoint has three relevant stages: |
|
|
| 1. A rank-16 LoRA continuation mixed 8,000 audited local records with 8,000 |
| deterministic Kev replay records for one pass. |
| 2. A hard-label full-weight BF16 continuation ran one pass over 11,320 records |
| / 16,279 questions using six A100-SXM4-80GB GPUs, FSDP2, seed `92323`, and |
| LR/head LR `1e-6`. |
| 3. R2 froze six WiSE-FT candidates and selected a 75% local parent / 25% |
| hard-final interpolation using DEVELOPMENT and CALIBRATION only. This stage |
| adds no optimizer update or new label. |
|
|
| A matched Qwen3.5-27B teacher research arm was trained during R1, but the |
| released checkpoint uses the hard-label arm, not that teacher arm. Historical |
| K0 data did contain some local-27B soft targets; those response rows are not |
| redistributed and are disclosed as a byte-reproduction gap. |
|
|
| See `TRAINING.md` and `DATA.md` in this repository for the full round ledger, |
| data counts, source revisions, licenses, negative results, and hardware. |
|
|
| ## Quality evidence |
|
|
| The independently opened capability-v1 confirmation contains 1,200 records / |
| 1,920 questions. On the 1,800-question clean paired population: |
|
|
| | Model | Accuracy | Multiclass Brier | Coverage at <=5% empirical error | |
| |---|---:|---:|---:| |
| | OpenJev 9B | 0.93611 | 0.08726 | 0.97611 | |
| | K0 parent | 0.94556 | 0.07848 | 0.98944 | |
| | Raw Qwen3.5-9B scoring adapter | 0.81333 | 0.28569 | 0.30889 | |
|
|
| These are synthetic confirmation results, not proof of general intelligence or |
| Jev parity. A later corrected DEVELOPMENT-only comparison on true released |
| Kev-9B found OpenJev / Kev accuracy `0.757624 / 0.738898` and Brier |
| `0.328546 / 0.355000`; it is not a locked test. |
|
|
| ## Serving measurements |
|
|
| The same multilingual three-question Go request was measured after two warmups |
| and 20 requests at each concurrency. Every one of 160 measured CPU/GPU requests |
| returned HTTP 200. |
|
|
| | Device | c1 throughput / p50 | c8 throughput / p50 | Peak model memory | |
| |---|---:|---:|---:| |
| | 16-core Xeon CPU | 0.5699 req/s / 1,692.7 ms | 0.5037 req/s / 15,678.8 ms | 13.7 GiB RSS | |
| | A100-SXM4-80GB | 10.7123 req/s / 85.7 ms | 20.0055 req/s / 342.7 ms | 23.6 GiB device | |
|
|
| CPU is appropriate for offline, fallback, or very low-QPS use. A GPU is |
| recommended for interactive concurrent serving. Discrete CPU/GPU decisions |
| matched; probability bits are not promised identical across devices or dynamic |
| batches. |
|
|
| ## Data and reproducibility |
|
|
| OpenJev publishes source builders, pinned revisions, hashes, roles, and |
| attribution. Original deterministic fixtures can be released under their |
| declared terms. Third-party rows that require per-origin review, contain a |
| benchmark canary, or have share-alike/attribution constraints remain |
| download/build-recipe-only. Jev responses and private teacher-response ledgers |
| are excluded. |
|
|
| That means the model is open and independently verifiable, while exact |
| from-scratch byte reproduction still has a documented gap around historical K0 |
| rows/targets. See `DATA.md`; do not infer that all upstream dataset rows are |
| owned by OpenJev. |
|
|
| ## Limitations |
|
|
| - No fresh Jev baseline and no parity claim. |
| - The model inherits biases and factual limitations from Qwen, Kev, and the |
| listed training sources. |
| - Numeric/date/business-day and long-context joins remain known weak areas in |
| consumed reporting-only diagnostics; those rows were not mined for training. |
| - Concurrent GPU batching can change rounded probabilities by up to 0.0018 in |
| the measured fixture without changing the discrete decision. |
| - Active GPU compute is not preempted on client cancellation; queued work can |
| be cancelled cooperatively. |
| - The API is close to, but not certified identical to, every hosted Jev error, |
| usage, refusal, and rate-limit behavior. |
|
|
| ## License and attribution |
|
|
| OpenJev code and released weights are provided under Apache-2.0. This model is |
| derived from Apache-2.0 Qwen and Kev assets. See `LICENSE` and `NOTICE`. |
| “Jev” and “TypeSafe” are used only to describe compatibility and the research |
| target; this project is not affiliated with or endorsed by TypeSafe. |
|
|
