Дообучение медицинской ASR-модели Parakeet

★ 7.6 · content

digital-health-clinical-asr-finetune is a Claude Code skill that implements Stage 4 of the Clinical ASR Flywheel: it runs stock NeMo SFT on the Parakeet TDT v2 base model inside the nvcr.io/nvidia/nemo:25.11.01 container whenever the priority-category KER exceeds 0.3, then performs an offline cycle N+1 re-evaluation to confirm the loop has closed. On the reference 39-row manifest, KER dropped from 0.513 to 0.128 (−75% relative) in 3 epochs — drug names fell from 0.857 to 0.214, conditions and procedures reached 0.000. The skill uses Read, Write, Bash, and Skill tools, requires a CUDA host with at least 16 GB VRAM (24 GB recommended) or a Brev L40S instance, and needs an NVIDIA_API_KEY for the eval round-trip. Healthcare teams fine-tuning clinical ASR on specialized medical vocabulary will use the resulting .nemo file with riva-asr-custom for production NIM deployment.