OAK-Decision-parallel-decision-qwen3.5-0.8b

OAK-Decision · 并行决策 / Parallel Decision — 实验 1 / Experiment 1

一次编码 State 前缀，多个问题分支彼此隔离，pointer head 在一次前向计算中为给定选项打分；无自回归生成。 / One shared State prefix, isolated question branches, and option scores from one forward pass; no autoregressive generation.

中文

模型与原理

本模型基于 Qwen3.5-0.8B。State 与多个「问题—选项」组被打包到同一次前向计算中：问题读取共享 State，但不能读取其他问题；DeltaNet 分支状态从共同前缀分叉。Pointer head 读取各选项的隐藏状态并输出 logits。这里发布的是判别模型，不能用 AutoModelForCausalLM.generate() 代替仓库内的推理代码。LoRA rank 16、alpha 32、dropout 0.05；pointer head 维度 256。

使用

pip install -r requirements.txt
python inference.py --record example_record.json --device cuda
# 本地基座权重：附加 --base-model /path/to/Qwen3.5-0.8B

JSON 输入是字符串 state 和包含 id、question、options 的 questions 列表。输出只对调用方给出的选项生成 softmax 分数，未经概率校准。需另行取得 Qwen 基座权重并使用本仓库的自定义 src/ 代码。example_record.json 只演示接口，并非测试样本。

评估与边界

训练任务采用中文合成的儿童照护／家庭出行记录。在留出的合成测试集中，IID 2,016 个问题准确率 92.91%，OOD 2,185 个问题准确率 90.62%；选项顺序置换一致率 94.8%。这些数字不代表真实座舱性能，数据集未随模型发布。模型无法发现遗漏选项，也不会调用车辆 API；尚未证明其能可靠处理真实 ASR 噪声、未知服务或安全关键决策。

授权与署名： 模型权重及原创推理代码 © 2026 OpenSparX（上海塞伯火种人工智能有限公司），Apache-2.0。基座模型的署名与使用条款见 Qwen/Qwen3.5-0.8B。仓库不含私有凭据或训练数据。

English

Developer: OpenSparX (上海塞伯火种人工智能有限公司). License: Apache-2.0. This release contains the floating-point LoRA adapter, a separately trained pointer decision head, and the Python inference implementation. It does not contain the Qwen base weights; download Qwen3.5-0.8B separately. The architecture is a research implementation of shared-State, isolated-question, prefill-only decision scoring. It does not generate text.

Quick start

pip install -r requirements.txt
python inference.py --record example_record.json --device cuda
# Offline base-model copy: add --base-model /path/to/Qwen3.5-0.8B

inference.py accepts a JSON record with state (string) and questions (a list containing id, question, and options). It prints one softmax score distribution per question. Scores should not be treated as calibrated probabilities. The reference implementation loads the Qwen text backbone, applies the LoRA adapter, and executes the custom pointer head; AutoModelForCausalLM.generate() is not the inference path. For reproducible results use the pinned dependency versions and FP32 weights in this release.

Training and evaluation

The model was trained with LoRA rank 16, alpha 32 and dropout 0.05, plus a 256-dimensional pointer head. The source task used synthetic Chinese child-care/family-trip State/Question records. The held-out synthetic tests contain 2,016 IID and 2,185 OOD questions; per-question answer accuracy was 92.91% IID / 90.62% OOD. Option-order permutation consistency was 94.8%. These figures characterize the project’s synthetic benchmark, not real-world cabin performance. The benchmark dataset is not included in this model repository.

Scope and limitations

• The custom src/ code is required; the adapter alone cannot perform JEV scoring.
• The output is a choice score over caller-provided options. It cannot discover options that were omitted, and it does not call a vehicle API.
• The model was trained and evaluated on synthetic data. It has not been established as robust to real ASR noise, unseen services or safety-critical decisions.
• The public example_record.json demonstrates the interface, not a scored benchmark case.

Model weights and original JEV inference code: © 2026 OpenSparX, Apache-2.0. Base-model attribution and terms: Qwen/Qwen3.5-0.8B. No private service credentials or training data are included.