AudioJev — direct audio decisions with order-calibrated probabilities
Upload or record a clip, ask a question, and list the answer options. AudioJev returns a probability for every option in a single forward pass — it never generates text, it only scores the listed candidates.
Examples
| Decision type | Question / instruction | Options, one per line (choice) | Score levels, worst to best (score mode) | Audio clip |
|---|
Notes
- choice scores 2–36 options; the model was trained on alphabetically sorted option lists, so distinct non-relative options are re-sorted the same way before scoring (use 'key=description' to control the displayed key).
- noul answers a true/false statement about the audio.
- score spreads probability over ordered levels and reports their expectation.
- Confidence is distribution concentration (1 − normalized entropy), not the probability of being correct.
- Audio is capped at 20 MB and 300 s; best under 30 s. Weights: shlv/AudioJev (Qwen Research License). Example clips are CC0 / public-domain audio from Wikimedia Commons.