Imported from nvidia/skills (skills/digital-health-clinical-asr-finetune/SKILL.md) via skills.sh. Install upstream with npx skills add nvidia/skills --skill digital-health-clinical-asr-finetune. Copyright stays with the author (Apache-2.0).

Clinical ASR Flywheel — Stage 4 (Fine-tune)

⚠ Agent: read this entire SKILL.md before answering. The Critical-workflow-rules section, the base-model table (§4c), the stock-NeMo-SFT recipe (§4d), and the cycle-N+1 decision table (§4e) are all load-bearing — the do-not-SFT bases and broken-adapter warnings live there.

Agent: this file is self-contained. The Stage 4 gate criteria, base-model recommendation, hyperparameter table, container invocation pattern, and cycle-N+1 decision table are all below. Do not run file-discovery commands or open references/stage4-finetune.md before answering methodology questions — the reference is deep-dive material, not required reading. Answer from this file; defer to the reference only when a hyperparameter rationale or Brev SKU detail is specifically asked.

You are the adapt-and-measure stage. The user arrives from /digital-health-clinical-asr-eval with a manifest, a baseline KER number, and the decision-tree's recommendation that fine-tuning is worth the GPU time. You run stock NeMo SFT, do an offline cycle N+1 re-eval to measure that the loop closed, and optionally hand the resulting .nemo to /riva-asr-custom for production serving.

The cycle KER from offline eval is the measurement that closes the loop. Riva NIM deploy validates serving (latency, streaming, scale), not model quality.

Empirically verified on the reference manifest (39 rows, Parakeet TDT v2): Baseline KER 0.513 → after 3 epochs of stock SFT: 0.128 (-75% relative). Drug names: 0.857 → 0.214. Conditions: 0.500 → 0.000. Procedures: 0.250 → 0.000.

Critical workflow rules (apply on every activation)

Surface these facts in any response, even if the user asks a narrow question:

1. Read this entire SKILL.md before answering. The base-model selection table, hyperparameter values, and the cycle-N+1 decision table are below — they are the load-bearing parts.
2. Verified result — Parakeet TDT v2 with the recipe in §4c achieves KER 0.513 → 0.128 (−75% relative) in 3 epochs on the reference manifest. Cite this when the user asks whether SFT will help.
3. Recipe is /opt/NeMo/examples/asr/speech_to_text_finetune.py inside nvcr.io/nvidia/nemo:25.11.01. Stock script, no patches, no custom adapter logic. The adapter-mixin path is broken on TDT/RNNT decoders (72 NaN tensors at any LR) — do not propose it.
4. Recommended base is nvidia/parakeet-tdt-0.6b-v2. The full base-model table is in §4c.
5. Do NOT fine-tune nvidia/nemotron-speech-streaming-en-0.6b. The streaming NVCF function's SFT path is broken (UNK collapse on validation after step 1). For streaming serving at deploy time, Riva chunks a non-streaming base just fine. Warn the user proactively if they propose it.
6. Gate the recommendation. Stage 4 only fires when priority-category KER > 0.3 and manifest has ≥ 100 rows (≥ 5 per priority category). Below those thresholds, route back to /digital-health-clinical-asr-build to grow the manifest first.

Purpose

Run stock NeMo SFT (no custom adapter logic, no patches) in nvcr.io/nvidia/nemo:25.11.01 against a term-aware row-disjoint train/val split, produce a .nemo model, and re-eval offline as cycle N+1. Decide based on the cycle-N → cycle-N+1 KER delta whether to keep the model, grow the manifest, or accept that fine-tuning didn't help. Optionally hand the .nemo to /riva-asr-custom for NIM deploy.

When to use this skill

Activate on user phrases like:

• "Fine-tune ASR on my clinical vocabulary"
• "Improve ASR on medication names"
• "We have a KER of 0.4, can we fine-tune?"
• "Run SFT on my Parakeet TDT base"
• "Train a clinical ASR adapter"
• "Compare cycle 1 vs cycle 2 KER"
• "Deploy my fine-tuned model as a NIM" (this skill prepares the .nemo and routes to /riva-asr-custom for the deploy)

Do not activate when:

• The user hasn't scored a baseline yet → /digital-health-clinical-asr-eval
• The user doesn't have a manifest → /digital-health-clinical-asr-build
• The user wants generic word boosting / LM fusion (not SFT) → /finetune-asr
• The user has a .nemo and only wants to deploy → /riva-asr-custom

Prerequisites

• A cycle-N manifest + cycle-N eval result from /digital-health-clinical-asr-eval. The priority-category KER must be > 0.3 (Stage 4 gate). The manifest should have ≥ 100 rows total, and ≥ 5 rows per priority entity_category, for a believable post-tune signal.
• A CUDA host — 24 GB VRAM is comfortable for Parakeet TDT 0.6B at batch_size=4 with bf16-mixed; 16 GB works with smaller batch. No local GPU? Use Brev — recommended SKU is L40S 48 GB.
• The NeMo container: nvcr.io/nvidia/nemo:25.11.01. Pull once: docker pull nvcr.io/nvidia/nemo:25.11.01.
• NVIDIA Container Toolkit + Docker — covered by /riva-nim-setup if not already installed.
• A train/val split stratified by entity_category (recipe sketch in Step 4b below).
• /riva-asr-custom installed if you intend to deploy. Pure-research SFT runs without it.

Instructions

4a. Provision a GPU host (skip if you already have one)

Stage 4 needs a CUDA host with ≥ 16 GB VRAM (24 GB comfortable). If you have a local one that fits, skip this section. If not, use Brev — NVIDIA's per-second-billed GPU host service. Recommended SKU: L40S 48 GB.

Cost disclosure — surface this to the user before any brev create. L40S 48 GB runs $1.50/hr at time of writing; a 3-epoch SFT run on a 100-row manifest finishes in 15–30 minutes ($0.40–$0.75 of compute). The real risk is forgetting to stop the instance — overnight idle on L40S is ~$36, a week of idle is ~$250. Mitigations: (a) always wrap the workflow in a script that ends with brev stop; (b) set a calendar reminder when you start; (c) brev delete instead of brev stop if you don't need to keep the disk (stop keeps disk at $0.10/GB-month — 200 GB ≈ $20/month of latent cost). Confirm the user accepts the per-hour cost shape and the idle risk before spinning anything up.

Full setup walkthrough — CLI install (download-then-run, not curl-pipe), SKU choice, disk sizing, SSH config — is in references/stage4-finetune.md (§Brev provisioning).

Short happy-path once the CLI is installed. Do not run brev create until the user has explicitly typed YES at the confirmation prompt below — the gate is mandatory, not advisory, because everything after it bills against the user's account by the second:

brev login                                  # browser auth

# Mandatory cost-confirmation gate — do NOT skip or auto-answer this.
echo "About to provision: digital-health-clinical-asr-sft on L40S 48 GB."
echo "Cost shape: ~\$1.50/hr while running; ~\$36/night if left idle; ~\$20/mo disk if you 'stop' instead of 'delete'."
read -rp "Type YES to provision (anything else cancels): " confirm
[ "$confirm" = "YES" ] || { echo "Cancelled — no GPU instance was created."; exit 1; }

brev create digital-health-clinical-asr-sft \
  --gpu l40s:1 --image ubuntu-22-04-cuda-12-4 --disk 200gi
brev ssh-config                             # writes ~/.ssh/config entries
rsync -avz ./cycle1/ digital-health-clinical-asr-sft:~/cycle1/
brev shell digital-health-clinical-asr-sft            # drops into the instance
nvidia-smi                                  # confirm GPU
docker pull nvcr.io/nvidia/nemo:25.11.01    # ~12 GB, once per instance

When done, always halt billing: brev stop digital-health-clinical-asr-sft (keeps disk) or brev delete digital-health-clinical-asr-sft (frees it). For path rewriting laptop → Brev → NeMo container, see references/container-paths.md.

4b. Term-aware train/val split

Row-disjoint, stratified by entity_category, default val fraction 0.2.

The same term may appear on both sides via different rows (different voice, context, noise). That's expected and desirable — it measures acoustic + contextual robustness on the trained vocabulary, which is the standard ASR adaptation metric.

Singleton categories (one row total) get forced to train with a warning. If any priority category has < 5 rows, bail to /digital-health-clinical-asr-build — held-out validation will be too noisy to attribute movement.

Sketch:

# After loading manifest.jsonl into a list of dicts `rows`:
from collections import defaultdict
import random
random.seed(42)

by_cat = defaultdict(list)
for r in rows:
    by_cat[r["entity_category"]].append(r)

train, val = [], []
for cat, cat_rows in by_cat.items():
    random.shuffle(cat_rows)
    if len(cat_rows) < 2:
        train.extend(cat_rows)
        print(f"warning: singleton category {cat}, forced to train")
        continue
    n_val = max(1, int(0.2 * len(cat_rows)))
    val.extend(cat_rows[:n_val])
    train.extend(cat_rows[n_val:])

Write train.jsonl and validation.jsonl alongside the manifest. These are the inputs to speech_to_text_finetune.py.

4c. Choose the base model

Other Parakeet/Conformer bases (1.1B, CTC, RNNT, stt_en_conformer_ctc_large) + decoder → NIM container mapping: references/stage4-finetune.md. If the user asks to fine-tune Nemotron Speech Streaming, warn about the collapse and recommend Parakeet TDT v2.

4d. Stock NeMo SFT

In the NeMo container, invoke /opt/NeMo/examples/asr/speech_to_text_finetune.py directly. No custom adapter logic. No patches. The stock NeMo SFT script is the verified working recipe.

Hyperparameters (verified on Parakeet TDT v2, 39-row manifest):

init_from_pretrained_model: nvidia/parakeet-tdt-0.6b-v2
precision:                  bf16-mixed       # required for TDT numerical stability
lr:                         3e-4             # CosineAnnealing schedule
warmup_steps:               5                # tiny manifest; bump to 500 at production scale
epochs:                     3                # smoke; 10-30 for production
batch_size:                 4                # fits 16 GB VRAM; raise to 16 on L40S 48 GB
gradient_clip_val:          1.0              # defensive

Container invocation: docker run --gpus all --rm -it -v "$PWD:/workspace" nvcr.io/nvidia/nemo:25.11.01 python /opt/NeMo/examples/asr/speech_to_text_finetune.py with model.train_ds.manifest_filepath=/workspace/train.jsonl, model.validation_ds.manifest_filepath=/workspace/validation.jsonl, init_from_pretrained_model=nvidia/parakeet-tdt-0.6b-v2, and the hyperparameter overrides from the table above. Full docker-run line with config-path / config-name flags: references/stage4-finetune.md §Container invocation.

Manifest paths inside the container. Host paths (e.g. $HOME/…) don't resolve in /workspace. Rewrite snippet: references/container-paths.md.

The training run writes adapted_model.nemo and a training_run_info.json summary. Both go into a per-cycle subdirectory of the user's choice (e.g. cycle<N>/models/<run>/; the layout doesn't matter as long as it's consistent across cycles).

4e. Offline cycle N+1 eval — close the loop

Re-transcribe the cycle's audio with the fine-tuned .nemo using NeMo's offline transcribe(). No Riva needed — this is measurement, not serving. NeMo's offline path runs the same encoder + decoder graph the Riva NIM eventually serves.

Sketch:

import nemo.collections.asr as nemo_asr
model = nemo_asr.models.ASRModel.restore_from("adapted_model.nemo")
hyps = model.transcribe(["audio/row1.wav", "audio/row2.wav", ...])

Score the same four metrics (WER/CER/KER/SER) and the same five-section leaderboard the eval skill produces. Write them as leaderboard_cycle<N+1>.md. Compare against leaderboard_cycle<N>.md.

Decision table — cycle-N+1 vs cycle-N:

4f. (Optional) Deploy as a Riva NIM

Hand the .nemo to /riva-asr-custom. Pass the source architecture explicitly — /riva-asr-custom can't reliably detect CTC vs RNNT vs TDT from the .nemo alone, and the wrong NIM container produces a broken RMIR with no clear error:

After deploy: re-run /digital-health-clinical-asr-eval against the new endpoint (ASR_ENDPOINT=localhost:50051) to validate that production-serving numbers match offline numbers. Any divergence is in Riva preprocessing or riva-build flags, not the model. Route to /riva-asr-custom.

Examples

Scenario A — gate met. User: "Drug KER 0.42, 130 rows. SFT?" → Yes (gate cleared). parakeet-tdt-0.6b-v2 (verified 0.513 → 0.128). No local GPU? Step 4a (Brev) → 4b (split) → 4d (stock SFT) → 4e (offline re-eval). If cycle-2 drug KER drops ≥ 20% relative, keep the .nemo; otherwise back to /digital-health-clinical-asr-build.

Scenario B — Nemotron Streaming. User: "SFT nvidia/nemotron-speech-streaming-en-0.6b?" → No (UNK collapse). Substitute parakeet-tdt-0.6b-v2. Riva chunks non-streaming bases for streaming serving — base doesn't need to be streaming-native.

Scenario C — cycle 2 KER unchanged. User: "KER barely moved." → Back to /digital-health-clinical-asr-build. Signal density beats LR sweeps. If magpie_g2p rows are bad but merriam-webster rows are good, the gap is pronunciation-coverage — /digital-health-clinical-asr-build Step 2d.

Artifacts produced

• train.jsonl, validation.jsonl — term-aware split (Step 4b)
• adapted_model.nemo — fine-tuned model (Step 4d)
• training_run_info.json — hyperparameters, dataset stats, end-of-train metrics
• offline_hyps.jsonl — cycle-N+1 transcription hypotheses (Step 4e)
• leaderboard_cycle<N+1>.md — cycle-N+1 five-section leaderboard
• (optional, after Step 4f) a deployed NIM endpoint (delegated to /riva-asr-custom)

Troubleshooting

• Stage 4 training collapses to all-UNK after first step → you're on the cache-aware streaming RNNT base (nemotron-speech-streaming-en-0.6b). Route to nvidia/parakeet-tdt-0.6b-v2 (the recommended default) or nvidia/stt_en_conformer_ctc_large (legacy fallback). The streaming RNNT SFT path is broken; do not retry with different hyperparameters.
• Manifest paths don't resolve inside the NeMo container → host paths (e.g. $HOME/…) need rewriting to /workspace/…. See references/container-paths.md for the rewrite snippet.
• Cycle N+1 KER unchanged from cycle N → on parakeet-tdt-0.6b-v2 with the recipe above, this almost always means manifest signal density is too low. Grow the manifest first; don't sweep LR. (If you're on an older adapter-style recipe instead of stock SFT, the adapter weights may not have moved off zero-init — switch to stock SFT.)
• Cycle N+1 KER got worse → overfit on a tiny manifest. Bail to /digital-health-clinical-asr-build and grow.
• Riva-served numbers diverge from offline numbers → the gap is in Riva preprocessing or riva-build flags, not the model. Route to /riva-asr-custom.
• bf16-mixed precision errors → some GPUs (older Turing, all Volta) don't support BF16. Drop to fp32 and reduce batch_size. Use fp16-mixed only if fp32 is too slow — fp16 with TDT decoders can produce NaN losses, so check loss curves early.
• OOM during training on 24 GB GPU → drop batch_size to 2, raise accumulate_grad_batches to 2 to keep the effective batch size constant.

Limitations

• Adapter-style SFT on TDT/RNNT decoders is broken. Empirically confirmed: an earlier LinearAdapter-mixin recipe produces 72 NaN tensors at any LR on TDT and RNNT decoders. Resolved by switching to NeMo's stock full-model SFT (speech_to_text_finetune.py) — which is what this skill recommends. Do not attempt adapter SFT on TDT/RNNT bases.
• Don't SFT nemotron-speech-streaming-en-0.6b. The streaming-only NVCF function's SFT path is unreliable (UNK collapse). For streaming serving at deploy time, Riva chunks a non-streaming base.
• Tiny manifests overfit fast. Below ~100 rows total or ~5 rows per priority category, cycle-N+1 numbers are noisy. Grow before trusting a small KER drop.
• English-only by default. The base-model table is en-US-specific. Other locales need a different base + a re-validated SFT recipe.
• No turn-key driver. The user writes their own training-driver layout — output paths, run naming, leaderboard re-rendering. The methodology and recipes transfer; exact cycle-1 numbers depend on the user's manifest.

Next steps

• Deploy the .nemo as a NIM: /riva-asr-custom (pass the source architecture explicitly).
• Grow the manifest for cycle N+2: /digital-health-clinical-asr-build.
• Re-score the cycle: /digital-health-clinical-asr-eval (against the new endpoint or the new .nemo directly).
• Lateral for word boosting / LM fusion / non-clinical SFT recipes: /finetune-asr.

References

• references/stage4-finetune.md — base-model selection table, hyperparameter rationale, decoder → NIM container mapping, decision tree comparing cycle-N+1 to cycle-N
• references/container-paths.md — host → /workspace/ path rewriting for cross-host manifest portability (laptop ↔ Brev ↔ NeMo container)