v1tavitavita/voxmind-medical-action-state-stress

VoxMind Medical Action-State Stress

This dataset contains a synthetic Chinese server-TTS benchmark for executable reliability in spoken medical front-desk tool agents.

The main split is Argument-Binding Stress: 140 cases targeting appointment IDs, dates, departments, patient relation, final-after-partial turns, and pending cancellation. It is designed to test whether a spoken tool agent selects the correct action, tool, arguments, and final state before side-effecting execution.

Files

• —data/tts_stress_v45_args_binding_manifest.json: benchmark manifest with expected action, tool, arguments, state metadata, and audio paths.
• —data/tts_stress_v45_args_binding_audio/*.wav: synthetic server-TTS WAV files referenced by the manifest.
• —runs/voxmind_medical_tts_stress_v45_args_binding.jsonl: SFT-style rows for VoxMind action-token formatting.
• —runs/stress_v45_args_binding_data_generation_public.json: generation and audit summary with local paths removed.
• —runs/stress_v45_args_binding_tts_resynthesis_public.json: public TTS synthesis report.

Data Quality

• —Cases: 140
• —Label audit errors: 0
• —Exact text overlap with prior training/history data: 0
• —Fixture-id overlap: 0
• —Manifest-oracle strict accuracy: 140/140
• —Pending-cancel oracle clear: 4/4

Intended Use

The dataset is intended for research on spoken tool-agent execution, especially action selection, tool binding, argument binding, confirmation state, and cancellation behavior.

It is not intended for:

• —clinical advice
• —diagnosis
• —hospital deployment
• —evaluating real patient speech
• —replacing clinical safety review

Privacy

The dataset is synthetic. It contains no real patient audio and no protected health information.

Baseline Result

With the released final adapter:

Related Artifacts

• —Code: https://github.com/vittta/e2e-voxmind-medical-agent
• —Adapter: https://huggingface.co/v1tavitavita/voxmind-medical-action-state-adapter

License

Apache-2.0.