OAK-Decision-encoder-cabin-rbt3

OAK-Decision · 座舱 Encoder-JEV / Cabin Encoder-JEV — 浮点版 / Floating-Point Release

车辆状态与用户语句只编码一次；六个独立问题经 cross-attention 读取共享 State，再输出各自的选项分数。 / Vehicle State and utterance are encoded once; six isolated questions read the shared State through cross-attention and produce option scores.

中文

模型与原理

模型部分初始化自 hfl/rbt3。共享 State encoder 有 3 层，六个独立问题复用同一个 1 层 Branch encoder；每个分支通过 cross-attention 读取共享 State 的 K/V，而不读取其他问题。固定选项槽位经有效选项掩码后输出六组 logits。model.safetensors 是完整浮点模型张量，不是 QNN/HTP 编译文件。

使用

pip install -r requirements.txt
python inference.py --utterance '车里有点热' \
  --state-json '{"cabin_temp_c":29}' --device cuda

默认 State 还提供车速、电量、续航预估、目的地距离和座位占用。命令输出一组六选一路由判别及四个独立服务的是／否 softmax 分数；这些分数未经校准，不会调用车辆 API。config.json 固定训练时维度：State 长度 128、六个问题、分支长度 96、六个选项槽、hidden size 768、State 3 层、分支 1 层。

评估与边界

中文措辞候选由 DeepSeek 生成并经人工复核，标签由项目确定性规则产生；合成语料不随模型发布。1,101 条训练和 284 条验证记录采用模板分离措辞；第 4 epoch 选出的最佳 checkpoint 在验证集上六问题整条全对 282/284（99.30%）。另有 50 条脚本化演示集，FP32 模型单独整条全对 42/50。演示集后来被查看并用于开发可选 CPU 仲裁，因此 50/50 混合系统结果只代表已排练回归，不是模型单独准确率或独立泛化证据；仲裁层未发布。量化后继版本曾在 SA8397/Hexagon V81 上编译运行，但本仓库不含 QNN context 或高通 SDK 文件。State 是演示输入，尚未证明真实 ASR、缺失信号或安全关键场景的可靠性。

授权与署名： 模型权重与原创推理代码 © 2026 OpenSparX（上海塞伯火种人工智能有限公司），Apache-2.0。预训练初始化署名：hfl/rbt3，Apache-2.0。本仓库不含车辆 API 目录、服务密钥或用户数据。

English

Developer: OpenSparX (上海塞伯火种人工智能有限公司). License: Apache-2.0. This is the floating-point PyTorch/CUDA Encoder-JEV model, tokenizer, and inference code. It is initialized partly from hfl/rbt3 and contains a trained shared-State encoder, six isolated decision branches, and a custom option head. model.safetensors is a full set of model tensors, not a QNN or HTP binary.

Quick start

pip install -r requirements.txt
python inference.py --utterance '车里有点热' \
  --state-json '{"cabin_temp_c":29}' --device cuda

The default state supplies speed, cabin temperature, SOC, predicted range, destination distance, and seat occupancy. The command prints a six-way routing decision and four individual service yes/no softmax scores. These scores are uncalibrated and must not be presented as reliable probabilities. No vehicle API is called. config.json fixes the trained dimensions: S128, six questions, B96, six option slots, hidden size 768, three State layers and one branch layer.

Training and evaluation

Chinese wording candidates were generated with DeepSeek and manually reviewed; deterministic project rules produced the training labels. The release contains no synthetic training corpus. The resulting 1,101 train / 284 validation records use template-separated wording. Best checkpoint selection used the validation set at epoch 4, where six-question exact match was 282/284 (99.30%). On a separate 50-case scripted demonstration set, the FP32 model-only full-record match was 42/50. The scripted set was later inspected to develop optional CPU arbitration; a 50/50 hybrid result is therefore a rehearsed regression, not model-only accuracy or independent generalization. The arbitration layer is not included in this model release.

A quantized descendant was compiled and run on SA8397/Hexagon V81 as a separate engineering experiment. Its QNN context and Qualcomm SDK binaries are not included here. This repository is for CUDA/CPU floating-point inference only.

Scope and limitations

• The state fields are demonstration inputs, not authenticated live vehicle signals. The model does not execute car controls.
• Training and validation are synthetic; real ASR, unseen utterances, missing signals, and safety-critical behavior need independent evaluation.
• The six outputs share the State representation but are not an autoregressive sequence. A user-facing action policy needs explicit validation and calibrated thresholds.

Model weights and original inference code: © 2026 OpenSparX, Apache-2.0. Pretrained initialization attribution: hfl/rbt3, Apache-2.0. No QNN files, vehicle API catalogs, service secrets, or user data are included.