v1tavitavita/voxmind-medical-action-state-adapter

VoxMind Medical Action-State Adapter

This repository releases a LoRA adapter for a Chinese spoken medical front-desk tool-agent benchmark. The adapter targets an action-state execution interface for appointment creation, rescheduling, cancellation, read-only queries, callback requests, pending cancellation, and human handoff.

The adapter is trained for the local <med_action>{...}</med_action> contract used in the accompanying code release:

• Code: https://github.com/vittta/e2e-voxmind-medical-agent
• Dataset: https://huggingface.co/datasets/v1tavitavita/voxmind-medical-action-state-stress

Model Type

• Base model: leungtianle/VoxMind
• Released artifact: PEFT/LoRA adapter only
• Language: Chinese
• Data: synthetic medical front-desk server-TTS data
• License: Apache-2.0

This repository does not include the VoxMind base model. Users must obtain and prepare the base model separately according to the VoxMind release.

Intended Use

This adapter is intended for research on executable reliability in spoken tool agents. It should be used with the verifier, confirmation memory, and executor in the code repository.

It is not intended for:

• hospital deployment
• diagnosis or treatment advice
• real patient audio processing
• clinical decision support
• live telephony use without additional validation

Evaluation Summary

On the held-out Argument-Binding Stress subset:

System	Strict	Action-only	Tool mismatch	Args mismatch	Pending clear
Previous adapter	12/140	135/140	9	114	4/4
This adapter	113/140	138/140	4	21	4/4

Safety/state metrics for this adapter on the main subset:

• false_commit=0
• unsafe_executed_forbidden=0
• pending_clear=4/4

The evaluation is controlled server-TTS plus offline replay. It does not validate live WebRTC timing, VAD, TTS-stop timing, real patient robustness, or clinical deployment.

Loading

Use this adapter with the VoxMind base model and the project runtime:

python3 -m src.pipeline \
  --manifest data/tts_stress_v45_args_binding_manifest.json \
  --model-provider voxmind-lora \
  --voxmind-repo /path/to/VoxMind \
  --voxmind-model /path/to/VoxMind-base \
  --voxmind-adapter /path/to/this-adapter \
  --device cuda \
  --verifier-aware-decoding

Limitations

• Synthetic Chinese server-TTS data only.
• No real patient audio and no protected health information.
• Offline replay evaluation only.
• Narrow medical administrative workflow, not general medical reasoning.
• Requires the external VoxMind base model.

Citation

If you use this adapter, cite the accompanying repository and the VoxMind base model.