Medivoice baseline-omni-gepard-v1 (merged)
LoRA fine-tune of google/gemma-4-12B-it (revision 707f0a3b8a3c7ad586ed01e27eafbad8a27dd0f7) merged into the full model, from run
baseline-omni-gepard-v1, best checkpoint 766. The model takes a practice's historical system
instructions, the conversation so far and the caller's audio turn, and answers with MediVoice
DSL function calls followed by exactly one transcript("...") annotation of the current utterance.
Text-only tool continuations (tool result -> next calls) carry no transcript.
Output contract appended to the system prompt during training:
The final user input is audio. Respond with the appropriate MediVoice API function calls first, followed by exactly one transcript("...") annotation containing the current utterance. transcript is a training annotation, not an executable practice API function. Do not transcribe earlier turns.
Serving code must parse and strip the transcript(...) annotation before dispatching practice calls.
Training
| Adapter | LoRA r=16, alpha=32, dropout 0.05 on q/k/v/o/gate/up/down of the language model (328 modules); audio encoder frozen |
| Optimizer | AdamW lr 5e-5, wd 0.01, cosine schedule, 3% warmup, grad accumulation 16, 2 epochs = 766 steps, seed 1729 |
| Loss | target-only cross-entropy incl. the transcript annotation; max length 32768, no truncation |
| Data | 6,121 train sources (3,550 audio + 2,571 text continuations); 6,186 synthetic audio variants (OmniVoice, Gepard TTS) accepted by ASR agreement |
| Compute | 1x H200, 17.7 h |
Metrics (synthetic audio, session split; not real-call performance)
| Mean target-token loss / source | Audio | Text continuation |
|---|---|---|
| Base model, adapter disabled (validation) | 2.311 | 1.494 |
| Checkpoint 766 (validation) | 0.265 | 0.262 |
| Checkpoint 766 (test) | 0.272 | 0.193 |
Held-out greedy generation, source-balanced, rescored against historical labels: exact parsed-call agreement 19.1% (audio) / 34.8% (text); call-name sequence agreement 75.1% / 87.2%; transcript exact agreement 67.5%, mean WER 18.2%. Historical labels are not gold; see the training repo's curation docs.
Full details in training_metadata.json. Source: medivoice-training repo, branch mr-first.
Internal research artifact; not validated for production or clinical use.