Find your next open-source project

Search by name, purpose, or keyword — or paste a GitHub repository URL
All skills

digital-health-clinical-asr-finetune

Fine-tune Parakeet TDT v2 with stock NeMo SFT and re-evaluate KER offline.

  • Part of NVIDIA/skills
  • 1.9k installs
  • Machine Learning · Speech models

When to use it

You already have a scored baseline manifest with priority-category KER above 0.3 and need to know whether fine-tuning actually lowers it.

For example

You say

Use digital-health-clinical-asr-finetune on my 130-row clinical manifest; drug-name KER is 0.42 and I want to see if SFT brings it down.

You get
  • adapted_model.nemo
  • leaderboard_cycle2.md
  • KER delta decision

How it works

  • Splits the manifest row-disjoint and stratified by entity_category, defaulting to 20% validation.
  • Uses the stock speech_to_text_finetune.py script with bf16-mixed precision for TDT stability.
  • Warns against fine-tuning nemotron-speech-streaming-en-0.6b because its SFT path collapses.

Before installing

Needs a CUDA host (16 GB VRAM minimum), the NeMo container nvcr.io/nvidia/nemo:25.11.01, and the finetune-asr and riva-asr-custom skills installed alongside; NVIDIA_API_KEY is required for the offline cycle N+1 eval and any NIM deploy.