Initial private GGUF release (BF16, Q8_0, imatrix Q4_K_M)

.gitattributes

@@ -33,3 +33,7 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+APUS-OpenJev-v1-4B-BF16.gguf filter=lfs diff=lfs merge=lfs -text
+APUS-OpenJev-v1-4B-Q4_K_M.gguf filter=lfs diff=lfs merge=lfs -text
+APUS-OpenJev-v1-4B-Q8_0.gguf filter=lfs diff=lfs merge=lfs -text
+imatrix.gguf filter=lfs diff=lfs merge=lfs -text

APUS-OpenJev-v1-4B-BF16.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:843fbf2825172bd8b89b57752c6d7bcbc19c82fae6100c262aeed1f063bc06d7
+size 8424393568

APUS-OpenJev-v1-4B-Q4_K_M.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3e77f041b48a6b28081071dc4dca8b047c10acf0a4b9b9b48603cef6b06ee186
+size 2708804768

APUS-OpenJev-v1-4B-Q8_0.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5e57075a169a76de5150f5c5defd805525ce4df86ae8865f70f128e15e57416c
+size 4482403168

LICENSE

@@ -0,0 +1,202 @@
+
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright 2026 Alibaba Cloud
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

Modelfile

@@ -0,0 +1,14 @@
+# ollama create openjev-4b -f Modelfile   (change FROM to pick another quant)
+FROM ./APUS-OpenJev-v1-4B-Q8_0.gguf
+TEMPLATE """<|im_start|>user
+{{ .Prompt }}<|im_end|>
+<|im_start|>assistant
+<think>
+
+</think>
+
+"""
+PARAMETER temperature 0
+PARAMETER num_predict 1
+PARAMETER num_ctx 9216
+PARAMETER stop <|im_end|>

README.md

@@ -0,0 +1,72 @@
+---
+library_name: gguf
+license: apache-2.0
+base_model: apus-ailab/APUS-OpenJev-v1-4B
+base_model_relation: quantized
+pipeline_tag: text-generation
+language:
+- en
+- zh
+tags:
+- apus-openjev
+- decision-model
+- gguf
+- llama.cpp
+- ollama
+---
+
+# APUS-OpenJev-v1-4B-GGUF
+
+[English](README.md) | [中文](README.zh-CN.md) · [Source model](https://huggingface.co/apus-ailab/APUS-OpenJev-v1-4B) · [Collection](https://huggingface.co/collections/apus-ailab/apus-openjev-v1-6ab1ee888eb002fcdd3a2825) · [MLX (Apple Silicon)](https://huggingface.co/apus-ailab/APUS-OpenJev-v1-4B-MLX-8bit)
+
+GGUF conversions of [APUS-OpenJev-v1-4B](https://huggingface.co/apus-ailab/APUS-OpenJev-v1-4B) (revision `65797c526c27`) for **Ollama, llama.cpp and LM Studio** on Linux, Windows and macOS (Metal).
+
+OpenJev is a **decision model**: each request supplies a state, an instruction and 2–16 candidates; the model answers with one candidate label (A–P) and the application reads the distribution over those labels. It is not a chat model.
+
+## Files and parity
+
+Every file was scored on the [Frozen80 panel](https://huggingface.co/datasets/apus-ailab/APUS-OpenJev-Eval-Frozen80) with identical prompt tokens and compared with the HF BF16 release run by its own runtime (full depth, **66/80 · 82.50%**).
+
+| File | Size | Frozen80 | Decisions = HF BF16 | Max Δp vs HF BF16 |
+|---|---:|---:|---:|---:|
+| [Q8_0](APUS-OpenJev-v1-4B-Q8_0.gguf) | 4.2 GiB | 67/80 · 83.75% | 79/80 | 0.1288 |
+| [Q4_K_M](APUS-OpenJev-v1-4B-Q4_K_M.gguf) | 2.5 GiB | 67/80 · 83.75% | 77/80 | 0.8879 |
+| [BF16](APUS-OpenJev-v1-4B-BF16.gguf) | 7.8 GiB | 66/80 · 82.50% | 80/80 | 0.0780 |
+
+Q8_0 is the recommended default; Q4_K_M uses an importance matrix. Quantized scores are measured separately and do not inherit the BF16 result. Frozen80 is a reused development panel, not a blind benchmark. Per-file details: [evaluation/](evaluation/).
+
+## Ollama
+
+```bash
+ollama run hf.co/apus-ailab/APUS-OpenJev-v1-4B-GGUF:Q8_0
+```
+
+The repository ships the Ollama `template` (Qwen chat turn with thinking disabled) and `params` (`temperature 0`, `num_predict 1`). For a local import use the bundled [Modelfile](Modelfile). To get candidate probabilities, send the rendered prompt through the API:
+
+```bash
+python examples/openjev_local.py --backend ollama --model hf.co/apus-ailab/APUS-OpenJev-v1-4B-GGUF:Q8_0
+```
+
+Ollama returns at most 20 `top_logprobs` and cannot report named tokens, so the distribution is exact only when every candidate label is in the top 20 (Q8_0: **64/80** Frozen80 prompts); otherwise use the selected label or llama-server. 
+
+## llama.cpp (exact distribution)
+
+```bash
+llama-server -m APUS-OpenJev-v1-4B-Q8_0.gguf -c 9216 -ngl 999
+python examples/openjev_local.py --backend llama-server --url http://127.0.0.1:8080
+```
+
+`examples/openjev_local.py` renders prompts with [openjev_contracts.py](openjev_contracts.py), the same contract used in training.
+
+## Conversion
+
+- llama.cpp `b11118` (`e6ab7c1a4`), `convert_hf_to_gguf.py --no-mtp` (the merged release has no MTP weights).
+- Q4_K_M importance matrix: 448 training-course decisions, 64 per source, disjoint from Frozen80 ([details](evaluation/imatrix-calibration.json), [imatrix.gguf](imatrix.gguf)).
+- All 1-D tensors, including GDN `A_log` / `dt_bias` and norms, stay F32 in every file ([check](evaluation/tensor-check.json)).
+- Scope: full depth only (the 16/20-layer `low` exit is not available), text only (no vision tower), probabilities are **not calibrated**.
+
+## License
+
+Apache-2.0, inherited from the source model; see [LICENSE](LICENSE). Base model: [Qwen/Qwen3.5-4B](https://huggingface.co/Qwen/Qwen3.5-4B).
+
+**Authors:** gumpcheng ([xDAN2099](https://huggingface.co/xDAN2099)), zhangxu, [APUS AI-LAB](https://github.com/APUS-AI-Lab).

README.zh-CN.md

@@ -0,0 +1,72 @@
+---
+library_name: gguf
+license: apache-2.0
+base_model: apus-ailab/APUS-OpenJev-v1-4B
+base_model_relation: quantized
+pipeline_tag: text-generation
+language:
+- en
+- zh
+tags:
+- apus-openjev
+- decision-model
+- gguf
+- llama.cpp
+- ollama
+---
+
+# APUS-OpenJev-v1-4B-GGUF
+
+[English](README.md) | [中文](README.zh-CN.md) · [源模型](https://huggingface.co/apus-ailab/APUS-OpenJev-v1-4B) · [Collection](https://huggingface.co/collections/apus-ailab/apus-openjev-v1-6ab1ee888eb002fcdd3a2825) · [MLX（Apple Silicon）](https://huggingface.co/apus-ailab/APUS-OpenJev-v1-4B-MLX-8bit)
+
+[APUS-OpenJev-v1-4B](https://huggingface.co/apus-ailab/APUS-OpenJev-v1-4B)（revision `65797c526c27`）的 GGUF 版本，适用于 Linux / Windows / macOS（Metal）上的 **Ollama、llama.cpp、LM Studio**。
+
+OpenJev 是**决策模型**：每个请求给出状态、指令和 2–16 个候选，模型回答一个候选标签（A–P），应用读取这些标签上的概率分布。它不是聊天模型。
+
+## 文件与一致性
+
+每个文件都在 [Frozen80](https://huggingface.co/datasets/apus-ailab/APUS-OpenJev-Eval-Frozen80) 上用完全相同的 prompt token 评测，并与 HF BF16 发布版（其自带 runtime、完整深度，**66/80 · 82.50%**）对比。
+
+| 文件 | 大小 | Frozen80 | 与 HF BF16 决策一致 | 相对 HF BF16 最大 Δp |
+|---|---:|---:|---:|---:|
+| [Q8_0](APUS-OpenJev-v1-4B-Q8_0.gguf) | 4.2 GiB | 67/80 · 83.75% | 79/80 | 0.1288 |
+| [Q4_K_M](APUS-OpenJev-v1-4B-Q4_K_M.gguf) | 2.5 GiB | 67/80 · 83.75% | 77/80 | 0.8879 |
+| [BF16](APUS-OpenJev-v1-4B-BF16.gguf) | 7.8 GiB | 66/80 · 82.50% | 80/80 | 0.0780 |
+
+默认推荐 Q8_0；Q4_K_M 使用 importance matrix。量化成绩单独测量，不沿用 BF16 成绩。Frozen80 是复用的开发面板，不是盲测。明细见 [evaluation/](evaluation/)。
+
+## Ollama
+
+```bash
+ollama run hf.co/apus-ailab/APUS-OpenJev-v1-4B-GGUF:Q8_0
+```
+
+仓库自带 Ollama 的 `template`（关闭 thinking 的 Qwen 对话格式）和 `params`（`temperature 0`、`num_predict 1`）。本地导入可用 [Modelfile](Modelfile)。需要候选概率时，通过 API 发送渲染好的 prompt：
+
+```bash
+python examples/openjev_local.py --backend ollama --model hf.co/apus-ailab/APUS-OpenJev-v1-4B-GGUF:Q8_0
+```
+
+Ollama 最多返回 20 个 `top_logprobs`，且不能指定 token，因此只有全部候选标签都在前 20 时分布才是精确的（Q8_0：Frozen80 中 **64/80** 题）；否则请使用所选标签，或改用 llama-server。
+
+## llama.cpp（精确分布）
+
+```bash
+llama-server -m APUS-OpenJev-v1-4B-Q8_0.gguf -c 9216 -ngl 999
+python examples/openjev_local.py --backend llama-server --url http://127.0.0.1:8080
+```
+
+`examples/openjev_local.py` 使用 [openjev_contracts.py](openjev_contracts.py) 渲染 prompt，与训练时的格式一致。
+
+## 转换说明
+
+- llama.cpp `b11118`（`e6ab7c1a4`），`convert_hf_to_gguf.py --no-mtp`（merged 发布版不含 MTP 权重）。
+- Q4_K_M 的 importance matrix：训练集中 448 条决策，每个来源 64 条，与 Frozen80 无重叠（[明细](evaluation/imatrix-calibration.json)、[imatrix.gguf](imatrix.gguf)）。
+- 所有 1-D 张量（含 GDN `A_log` / `dt_bias` 和各类 norm）在每个文件中都保持 F32（[检查结果](evaluation/tensor-check.json)）。
+- 范围：仅完整深度（不提供 16/20 层 `low` 出口），仅文本（不含视觉塔），概率**未经校准**。
+
+## 许可
+
+Apache-2.0，继承自源模型，见 [LICENSE](LICENSE)。基座模型：[Qwen/Qwen3.5-4B](https://huggingface.co/Qwen/Qwen3.5-4B)。
+
+**作者：** gumpcheng（[xDAN2099](https://huggingface.co/xDAN2099)）、zhangxu、[APUS AI-LAB](https://github.com/APUS-AI-Lab)。

SHA256SUMS

@@ -0,0 +1,4 @@
+843fbf2825172bd8b89b57752c6d7bcbc19c82fae6100c262aeed1f063bc06d7  ./APUS-OpenJev-v1-4B-BF16.gguf
+3e77f041b48a6b28081071dc4dca8b047c10acf0a4b9b9b48603cef6b06ee186  ./APUS-OpenJev-v1-4B-Q4_K_M.gguf
+5e57075a169a76de5150f5c5defd805525ce4df86ae8865f70f128e15e57416c  ./APUS-OpenJev-v1-4B-Q8_0.gguf
+76b68bd161e5b866e00ff7ec184460ef44622efa33e7f61e54d30c603b731a36  ./imatrix.gguf

conversion-provenance.json

@@ -0,0 +1,23 @@
+{
+  "llama_cpp": {
+    "tag": "b11118",
+    "commit": "e6ab7c1a41054a888ada952eab4c886444c2f5ad"
+  },
+  "ollama": "0.34.3",
+  "mlx": "0.32.2",
+  "mlx_lm": "0.31.3",
+  "convert_env": {
+    "torch": "2.11.0+cpu",
+    "transformers": "4.57.6"
+  },
+  "reference_runtime": {
+    "torch": "2.8.0+cu128",
+    "transformers": "5.16.1",
+    "effort": "high"
+  },
+  "hardware": "NVIDIA RTX PRO 6000 Blackwell 96GB (Linux); Apple M5 24GB for Mac checks",
+  "date": "2026-09-23",
+  "source_repository": "apus-ailab/APUS-OpenJev-v1-4B",
+  "source_revision": "65797c526c27c4d24f564333779162cd4a64328e",
+  "base_model": "Qwen/Qwen3.5-4B"
+}

evaluation/imatrix-calibration.json

@@ -0,0 +1,19 @@
+{
+  "train_file": "/workspace/ms-swift-jev/data/registered-course-v1-20260920/decision.train.jsonl",
+  "train_sha256": "172b7427c4dcfa1270750142d2ad9ab6d42ee1350f532137329798a5af7b92c2",
+  "seed": 20260923,
+  "per_source": 64,
+  "records": 448,
+  "sources": {
+    "osunlp/Mind2Web": 64,
+    "oracle_only": 64,
+    "mnli": 64,
+    "sgd": 64,
+    "google-research-datasets/go_emotions": 64,
+    "nvidia/HelpSteer3": 64,
+    "boolq": 64
+  },
+  "frozen80_overlap": 0,
+  "output_sha256": "b4acbf8ecf6c38b789d809f5c23b44d8ea584a8731c425a497642c40b8782187",
+  "characters": 1297640
+}

evaluation/llamacpp-BF16.summary.json

@@ -0,0 +1,36 @@
+{
+  "variant": "4B-BF16",
+  "rows": 80,
+  "correct": 66,
+  "reference_correct": 66,
+  "decision_agreement": 80,
+  "changed_panel_indexes": [],
+  "max_abs_probability_delta": 0.07798188169586251,
+  "mean_abs_probability_delta": 0.0036633186839281073,
+  "all_candidates_in_top20": 64,
+  "reference_choice_in_top20": 80,
+  "min_candidate_mass_full_vocab": 0.8304116117153292,
+  "seconds": 10.42,
+  "by_family": {
+    "browser": {
+      "rows": 16,
+      "correct": 14
+    },
+    "hs3": {
+      "rows": 16,
+      "correct": 12
+    },
+    "boolq": {
+      "rows": 16,
+      "correct": 16
+    },
+    "mnli": {
+      "rows": 16,
+      "correct": 11
+    },
+    "score": {
+      "rows": 16,
+      "correct": 13
+    }
+  }
+}

evaluation/llamacpp-Q4_K_M.summary.json

@@ -0,0 +1,40 @@
+{
+  "variant": "4B-Q4_K_M",
+  "rows": 80,
+  "correct": 67,
+  "reference_correct": 66,
+  "decision_agreement": 77,
+  "changed_panel_indexes": [
+    59,
+    64,
+    72
+  ],
+  "max_abs_probability_delta": 0.8879217194689062,
+  "mean_abs_probability_delta": 0.03401718616064667,
+  "all_candidates_in_top20": 64,
+  "reference_choice_in_top20": 80,
+  "min_candidate_mass_full_vocab": 0.8305545665216829,
+  "seconds": 10.65,
+  "by_family": {
+    "browser": {
+      "rows": 16,
+      "correct": 14
+    },
+    "hs3": {
+      "rows": 16,
+      "correct": 12
+    },
+    "boolq": {
+      "rows": 16,
+      "correct": 16
+    },
+    "mnli": {
+      "rows": 16,
+      "correct": 10
+    },
+    "score": {
+      "rows": 16,
+      "correct": 15
+    }
+  }
+}

evaluation/llamacpp-Q8_0.summary.json

@@ -0,0 +1,38 @@
+{
+  "variant": "4B-Q8_0",
+  "rows": 80,
+  "correct": 67,
+  "reference_correct": 66,
+  "decision_agreement": 79,
+  "changed_panel_indexes": [
+    72
+  ],
+  "max_abs_probability_delta": 0.12883429532209645,
+  "mean_abs_probability_delta": 0.006993976078097623,
+  "all_candidates_in_top20": 64,
+  "reference_choice_in_top20": 80,
+  "min_candidate_mass_full_vocab": 0.7869114107711379,
+  "seconds": 12.91,
+  "by_family": {
+    "browser": {
+      "rows": 16,
+      "correct": 14
+    },
+    "hs3": {
+      "rows": 16,
+      "correct": 12
+    },
+    "boolq": {
+      "rows": 16,
+      "correct": 16
+    },
+    "mnli": {
+      "rows": 16,
+      "correct": 11
+    },
+    "score": {
+      "rows": 16,
+      "correct": 14
+    }
+  }
+}

evaluation/mlx-8bit.summary.json

@@ -0,0 +1,35 @@
+{
+  "variant": "4B-MLX-8bit",
+  "backend": "mlx 0.32.2 on Linux x86_64 (Device(gpu, 0))",
+  "rows": 80,
+  "correct": 66,
+  "reference_correct": 66,
+  "decision_agreement": 80,
+  "changed_panel_indexes": [],
+  "max_abs_probability_delta": 0.1201455295085907,
+  "mean_abs_probability_delta": 0.005359249900720897,
+  "peak_memory_gb": 6.09,
+  "seconds": 105.89,
+  "by_family": {
+    "browser": {
+      "rows": 16,
+      "correct": 14
+    },
+    "hs3": {
+      "rows": 16,
+      "correct": 12
+    },
+    "boolq": {
+      "rows": 16,
+      "correct": 16
+    },
+    "mnli": {
+      "rows": 16,
+      "correct": 11
+    },
+    "score": {
+      "rows": 16,
+      "correct": 13
+    }
+  }
+}

evaluation/ppl-crosscheck.summary.json

@@ -0,0 +1,15 @@
+{
+  "hf_bf16": [
+    22.1184,
+    9.8833,
+    7.1177
+  ],
+  "gguf_bf16_imatrix": [
+    22.0867,
+    9.9745,
+    7.1646
+  ],
+  "metric": "cumulative PPL over the second half of each 4096-token chunk of the imatrix calibration text (llama-imatrix convention), first 3 chunks",
+  "hf_script": "hf_chunk_ppl.py",
+  "note": "diagnostic cross-check of GGUF numerics against HF; high 9B values are a property of checkpoint-3000 on raw long text, reproduced by HF"
+}

evaluation/reference.summary.json

@@ -0,0 +1,6 @@
+{
+  "model_dir": "src/4B",
+  "effort": "high",
+  "rows": 80,
+  "correct": 66
+}

evaluation/tensor-check.json

@@ -0,0 +1,45 @@
+{
+  "APUS-OpenJev-v1-4B-BF16.gguf": {
+    "architecture": "qwen35",
+    "tensors": 426,
+    "tensor_types": {
+      "BF16": 249,
+      "F32": 177
+    },
+    "one_dim_tensors_checked": 153,
+    "gdn_decay_tensor_types": {
+      "F32": 48
+    },
+    "one_dim_non_f32": [],
+    "bytes": 8424393568
+  },
+  "APUS-OpenJev-v1-4B-Q8_0.gguf": {
+    "architecture": "qwen35",
+    "tensors": 426,
+    "tensor_types": {
+      "F32": 177,
+      "Q8_0": 249
+    },
+    "one_dim_tensors_checked": 153,
+    "gdn_decay_tensor_types": {
+      "F32": 48
+    },
+    "one_dim_non_f32": [],
+    "bytes": 4482403168
+  },
+  "APUS-OpenJev-v1-4B-Q4_K_M.gguf": {
+    "architecture": "qwen35",
+    "tensors": 426,
+    "tensor_types": {
+      "F32": 177,
+      "Q6_K": 33,
+      "Q4_K": 216
+    },
+    "one_dim_tensors_checked": 153,
+    "gdn_decay_tensor_types": {
+      "F32": 48
+    },
+    "one_dim_non_f32": [],
+    "bytes": 2708804768
+  }
+}

examples/openjev_local.py

@@ -0,0 +1,109 @@
+"""Call an APUS-OpenJev GGUF through Ollama or llama-server (standard library only).
+
+    python examples/openjev_local.py --backend ollama --model hf.co/apus-ailab/APUS-OpenJev-v1-4B-GGUF:Q8_0
+    python examples/openjev_local.py --backend llama-server --url http://127.0.0.1:8080
+
+The model scores caller-supplied candidates; it does not write JSON. The prompt is
+rendered by ``openjev_contracts.py`` (identical to the training contract) and wrapped
+in the no-thinking Qwen chat turn. Candidate labels are A..P, one token each.
+
+* llama-server returns the exact candidate distribution (full-vocabulary logprobs).
+* Ollama exposes at most 20 ``top_logprobs``. The distribution is exact when every
+  candidate label ranks in the top 20; otherwise only the selected label is reliable
+  and ``distribution_complete`` is False.
+"""
+
+import argparse
+import json
+import math
+import sys
+import urllib.request
+from pathlib import Path
+
+sys.path.insert(0, str(Path(__file__).resolve().parent.parent))
+from openjev_contracts import label_mapping, render_prompt  # noqa: E402
+
+CHAT = "<|im_start|>user\n{prompt}<|im_end|>\n<|im_start|>assistant\n<think>\n\n</think>\n\n"
+
+
+def _post(url, payload):
+    request = urllib.request.Request(url, json.dumps(payload).encode(), {"Content-Type": "application/json"})
+    with urllib.request.urlopen(request, timeout=600) as response:
+        return json.loads(response.read())
+
+
+def _normalize(labels, logprobs):
+    peak = max(logprobs[label] for label in labels)
+    weights = {label: math.exp(logprobs[label] - peak) for label in labels}
+    total = sum(weights.values())
+    return {label: weight / total for label, weight in weights.items()}
+
+
+def decide(request, backend="ollama", url=None, model=None):
+    mapping = label_mapping(request)  # validates the request; label -> candidate id
+    labels = list(mapping)
+    prompt = CHAT.format(prompt=render_prompt(request))
+    if backend == "ollama":
+        data = _post(
+            (url or "http://127.0.0.1:11434") + "/api/generate",
+            {
+                "model": model,
+                "prompt": prompt,
+                "raw": True,
+                "stream": False,
+                "logprobs": True,
+                "top_logprobs": 20,
+                "options": {"temperature": 0, "num_predict": 1, "num_ctx": 9216},
+            },
+        )
+        (position,) = data["logprobs"]
+        logprobs = {entry["token"]: entry["logprob"] for entry in position["top_logprobs"]}
+        selected = data["response"]
+    elif backend == "llama-server":
+        data = _post(
+            (url or "http://127.0.0.1:8080") + "/completion",
+            {"prompt": prompt, "n_predict": 1, "n_probs": 1024, "temperature": 0, "cache_prompt": True},
+        )
+        (position,) = data["completion_probabilities"]
+        logprobs = {entry["token"]: entry["logprob"] for entry in position["top_logprobs"]}
+        selected = position["token"]
+    else:
+        raise ValueError("backend must be ollama or llama-server")
+
+    complete = all(label in logprobs for label in labels)
+    result = {"type": request["primitive"], "distribution_complete": complete}
+    if complete:
+        probabilities = _normalize(labels, logprobs)
+        result["probabilities"] = {mapping[label]: p for label, p in probabilities.items()}
+        selected = max(probabilities, key=probabilities.get)
+    if selected not in mapping:
+        raise RuntimeError(f"model produced {selected!r}, not a candidate label")
+    if request["primitive"] == "choice":
+        result["choice"] = mapping[selected]
+    elif complete:
+        result["yes_probability"] = result["probabilities"]["yes"]
+    else:
+        result["answer"] = mapping[selected]
+    return result
+
+
+EXAMPLE = {
+    "id": "support-731",
+    "group_id": "support-731",
+    "primitive": "choice",
+    "state": "Order 731 was delivered. The customer confirms that the issue is resolved.",
+    "instructions": "Select the next support action.",
+    "criteria": [
+        {"id": "close_ticket", "description": "Close the ticket as resolved."},
+        {"id": "escalate", "description": "Escalate to a human agent."},
+        {"id": "refund", "description": "Issue a refund."},
+    ],
+}
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--backend", choices=("ollama", "llama-server"), default="ollama")
+    parser.add_argument("--url")
+    parser.add_argument("--model", help="Ollama model name")
+    args = parser.parse_args()
+    print(json.dumps(decide(EXAMPLE, args.backend, args.url, args.model), indent=2))

imatrix.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:76b68bd161e5b866e00ff7ec184460ef44622efa33e7f61e54d30c603b731a36
+size 3626528

openjev_contracts.py

@@ -0,0 +1,102 @@
+"""Small shared contract. Prompts use a strict whitelist of input fields."""
+
+import json
+import math
+
+PROMPT_VERSION = "jev.dynamic.prompt.v2"
+LABELS = tuple("ABCDEFGHIJKLMNOP")
+BINARY_CRITERIA = [
+    {"id": "yes", "description": "The stated proposition is true."},
+    {"id": "no", "description": "The stated proposition is false."},
+]
+
+
+def validate_request(record):
+    for key in ("id", "group_id", "state", "instructions"):
+        if not isinstance(record.get(key), str) or not record[key].strip():
+            raise ValueError(f"{key} must be a nonempty string")
+    if record.get("primitive") not in ("choice", "noul", "score_level"):
+        raise ValueError("unsupported primitive")
+    criteria = record.get("criteria")
+    if not isinstance(criteria, list) or not 2 <= len(criteria) <= len(LABELS):
+        raise ValueError("criteria must contain 2..16 candidates")
+    ids = []
+    for candidate in criteria:
+        if not isinstance(candidate, dict):
+            raise TypeError("candidate must be an object")
+        for key in ("id", "description"):
+            if not isinstance(candidate.get(key), str) or not candidate[key].strip():
+                raise ValueError(f"candidate {key} must be nonempty")
+        ids.append(candidate["id"])
+    if len(set(ids)) != len(ids):
+        raise ValueError("duplicate candidate ids")
+    if record["primitive"] != "choice" and criteria != BINARY_CRITERIA:
+        raise ValueError("noul and score_level require canonical yes/no criteria")
+
+
+def validate_record(record):
+    validate_request(record)
+    if record.get("gold") not in [c["id"] for c in record["criteria"]]:
+        raise ValueError("gold must be a candidate id")
+    if not isinstance(record.get("provenance"), dict):
+        raise TypeError("provenance must be an object")
+
+
+def label_mapping(record):
+    validate_request(record)
+    return dict(zip(LABELS, (c["id"] for c in record["criteria"])))
+
+
+def render_prompt_parts(record):
+    """Text prefix/suffix; callers MUST check tokenizer boundary equivalence."""
+    validate_request(record)
+    prefix = "Shared state:\n" + record["state"] + "\n\n"
+    task = {
+        "primitive": record["primitive"],
+        "instructions": record["instructions"],
+        "criteria": [
+            {"label": label, "description": candidate["description"]}
+            for label, candidate in zip(LABELS, record["criteria"])
+        ],
+    }
+    suffix = json.dumps(task, ensure_ascii=False, sort_keys=True)
+    suffix += (
+        "\nReturn only the selected letter: "
+        + ", ".join(LABELS[: len(record["criteria"])])
+        + ".\nAnswer:"
+    )
+    return prefix, suffix
+
+
+def render_prompt(record):
+    return "".join(render_prompt_parts(record))
+
+
+def to_messages(record):
+    validate_record(record)
+    inverse = {candidate: label for label, candidate in label_mapping(record).items()}
+    return {
+        "messages": [
+            {"role": "user", "content": render_prompt(record)},
+            {"role": "assistant", "content": inverse[record["gold"]]},
+        ]
+    }
+
+
+def format_response(record, probabilities):
+    """Map ordered candidate probabilities; score_level is NOT aggregate Score."""
+    mapping = label_mapping(record)
+    values = list(probabilities)
+    if len(values) != len(mapping) or any(
+        not math.isfinite(p) or p < 0 or p > 1 for p in values
+    ):
+        raise ValueError("invalid probabilities")
+    if not math.isclose(sum(values), 1, abs_tol=1e-5):
+        raise ValueError("probabilities must sum to one")
+    distribution = dict(zip(mapping.values(), values))
+    result = {"type": record["primitive"], "probabilities": distribution}
+    if record["primitive"] == "choice":
+        result["choice"] = max(distribution, key=distribution.get)
+    else:
+        result["yes_probability"] = distribution["yes"]
+    return result

params

@@ -0,0 +1,8 @@
+{
+  "temperature": 0,
+  "num_predict": 1,
+  "num_ctx": 9216,
+  "stop": [
+    "<|im_end|>"
+  ]
+}

template

@@ -0,0 +1,7 @@
+<|im_start|>user
+{{ .Prompt }}<|im_end|>
+<|im_start|>assistant
+<think>
+
+</think>
+