ser-probe

Overview ¶

Command ser-probe exercises the runtime/classify HF backend directly against an audio file or text string, so you can verify the inference pipeline works against your HF_TOKEN without running the full arena scenario (which depends on duplex providers behaving correctly).

Usage:

HF_TOKEN=... promptarena-ser-probe \
  -audio examples/voice-refund-demo/out/media/runs/<id>/<hash>.wav \
  -model superb/wav2vec2-base-superb-er \
  -label angry

HF_TOKEN=... promptarena-ser-probe \
  -text "I want a refund and I want it NOW" \
  -model SamLowe/roberta-base-go_emotions \
  -label anger

HF retired most audio-classification models from their free serverless tier in early 2026. If -audio returns "Model not supported by provider hf-inference", point -model at one that IS still warm or use a paid HF Inference Endpoint via -base-url.