[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill-nvidia-digital-health-clinical-asr-finetune":3,"mdc--tkoq7b-key":37,"related-org-nvidia-digital-health-clinical-asr-finetune":2793,"related-repo-nvidia-digital-health-clinical-asr-finetune":2951},{"slug":4,"name":4,"fn":5,"description":6,"org":7,"tags":11,"stars":26,"repoUrl":27,"updatedAt":28,"license":29,"forks":30,"topics":31,"repo":32,"sourceUrl":35,"mdContent":36},"digital-health-clinical-asr-finetune","fine-tune clinical ASR models","Stage 4 of the Clinical ASR Flywheel. Use when priority KER is above 0.3 to run stock NeMo SFT on Parakeet TDT v2 and offline cycle N+1 re-eval. NOT for generic word boosting (use \u002Ffinetune-asr).",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},"nvidia","NVIDIA","https:\u002F\u002Fpexgzepcugksgbtrxkhf.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Forg-logos\u002Fnvidia.png",[12,16,19,22,23],{"name":13,"slug":14,"type":15},"Healthcare","healthcare","tag",{"name":17,"slug":18,"type":15},"Machine Learning","machine-learning",{"name":20,"slug":21,"type":15},"Transcription","transcription",{"name":9,"slug":8,"type":15},{"name":24,"slug":25,"type":15},"Speech","speech",2473,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002Fskills","2026-07-14T05:27:24.3208","Apache-2.0",281,[],{"repoUrl":27,"stars":26,"forks":30,"topics":33,"description":34},[],"AI agent skills published by NVIDIA","https:\u002F\u002Fgithub.com\u002FNVIDIA\u002Fskills\u002Ftree\u002FHEAD\u002Fskills\u002Fdigital-health-clinical-asr-finetune","---\nname: \"digital-health-clinical-asr-finetune\"\ndescription: \"Stage 4 of the Clinical ASR Flywheel. Use when priority KER is above 0.3 to run stock NeMo SFT on Parakeet TDT v2 and offline cycle N+1 re-eval. NOT for generic word boosting (use \u002Ffinetune-asr).\"\nversion: \"1.0.0\"\nauthor: \"Ben Randoing \u003Cbrandoing@nvidia.com>\"\ntags:\n  - clinical-asr\n  - finetune\n  - sft\n  - nemo\n  - parakeet\n  - flywheel\ntools:\n  - Read\n  - Write\n  - Bash\n  - Skill\nlicense: Apache-2.0\ncompatibility: \"Requires a CUDA host (24 GB VRAM comfortable, 16 GB workable with batch_size=4), the NeMo container (nvcr.io\u002Fnvidia\u002Fnemo:25.11.01), and the finetune-asr + riva-asr-custom skills installed alongside this one. No local GPU? Use Brev. NVIDIA_API_KEY required for the offline cycle N+1 eval round-trip and for any NIM deploy.\"\nmetadata:\n  author: \"Ben Randoing \u003Cbrandoing@nvidia.com>\"\n  tags:\n    - clinical-asr\n    - flywheel\n    - finetune\n    - nemo-sft\n    - parakeet\n  team: healthcare-tme\n  domain: ai-ml\n  stage: 4\n  previous_skill: digital-health-clinical-asr-eval\n  next_skill: riva-asr-custom\n---\n\n\u003C!--\nSPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.\nSPDX-License-Identifier: Apache-2.0\n-->\n\n# Clinical ASR Flywheel — Stage 4 (Fine-tune)\n\n> **⚠ 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.\n\n> **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\u002Fstage4-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.\n\nYou are the **adapt-and-measure** stage. The user arrives from `\u002Fdigital-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 `\u002Friva-asr-custom` for production serving.\n\n**The cycle KER from offline eval is the measurement that closes the loop.** Riva NIM deploy validates serving (latency, streaming, scale), not model quality.\n\n> **Empirically verified on the reference manifest** (39 rows, Parakeet TDT v2):\n> Baseline KER **0.513** → after 3 epochs of stock SFT: **0.128** (-75% relative).\n> Drug names: 0.857 → 0.214. Conditions: 0.500 → 0.000. Procedures: 0.250 → 0.000.\n\n## Critical workflow rules (apply on every activation)\n\nSurface these facts in any response, even if the user asks a narrow question:\n\n1. **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.\n2. **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.\n3. **Recipe is `\u002Fopt\u002FNeMo\u002Fexamples\u002Fasr\u002Fspeech_to_text_finetune.py` inside `nvcr.io\u002Fnvidia\u002Fnemo:25.11.01`.** Stock script, no patches, no custom adapter logic. The adapter-mixin path is broken on TDT\u002FRNNT decoders (72 NaN tensors at any LR) — do not propose it.\n4. **Recommended base is `nvidia\u002Fparakeet-tdt-0.6b-v2`.** The full base-model table is in §4c.\n5. **Do NOT fine-tune `nvidia\u002Fnemotron-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.\n6. **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 `\u002Fdigital-health-clinical-asr-build` to grow the manifest first.\n\n## Purpose\n\nRun **stock NeMo SFT** (no custom adapter logic, no patches) in `nvcr.io\u002Fnvidia\u002Fnemo:25.11.01` against a term-aware row-disjoint train\u002Fval 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 `\u002Friva-asr-custom` for NIM deploy.\n\n## When to use this skill\n\nActivate on user phrases like:\n\n- \"Fine-tune ASR on my clinical vocabulary\"\n- \"Improve ASR on medication names\"\n- \"We have a KER of 0.4, can we fine-tune?\"\n- \"Run SFT on my Parakeet TDT base\"\n- \"Train a clinical ASR adapter\"\n- \"Compare cycle 1 vs cycle 2 KER\"\n- \"Deploy my fine-tuned model as a NIM\" *(this skill prepares the `.nemo` and routes to `\u002Friva-asr-custom` for the deploy)*\n\nDo **not** activate when:\n\n- The user hasn't scored a baseline yet → `\u002Fdigital-health-clinical-asr-eval`\n- The user doesn't have a manifest → `\u002Fdigital-health-clinical-asr-build`\n- The user wants generic word boosting \u002F LM fusion (not SFT) → `\u002Ffinetune-asr`\n- The user has a `.nemo` and only wants to deploy → `\u002Friva-asr-custom`\n\n## Prerequisites\n\n- **A cycle-N manifest + cycle-N eval result** from `\u002Fdigital-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.\n- **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.\n- **The NeMo container**: `nvcr.io\u002Fnvidia\u002Fnemo:25.11.01`. Pull once: `docker pull nvcr.io\u002Fnvidia\u002Fnemo:25.11.01`.\n- **NVIDIA Container Toolkit + Docker** — covered by `\u002Friva-nim-setup` if not already installed.\n- **A train\u002Fval split** stratified by `entity_category` (recipe sketch in Step 4b below).\n- **`\u002Friva-asr-custom`** installed if you intend to deploy. Pure-research SFT runs without it.\n\n## Instructions\n\n### 4a. Provision a GPU host (skip if you already have one)\n\nStage 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.\n\n**Cost disclosure — surface this to the user before any `brev create`.** L40S 48 GB runs ~$1.50\u002Fhr 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\u002FGB-month — 200 GB ≈ $20\u002Fmonth of latent cost). Confirm the user accepts the per-hour cost shape and the idle risk before spinning anything up.\n\nFull setup walkthrough — CLI install (download-then-run, not curl-pipe), SKU choice, disk sizing, SSH config — is in `references\u002Fstage4-finetune.md` (§Brev provisioning).\n\nShort 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:\n\n```bash\nbrev login                                  # browser auth\n\n# Mandatory cost-confirmation gate — do NOT skip or auto-answer this.\necho \"About to provision: digital-health-clinical-asr-sft on L40S 48 GB.\"\necho \"Cost shape: ~\\$1.50\u002Fhr while running; ~\\$36\u002Fnight if left idle; ~\\$20\u002Fmo disk if you 'stop' instead of 'delete'.\"\nread -rp \"Type YES to provision (anything else cancels): \" confirm\n[ \"$confirm\" = \"YES\" ] || { echo \"Cancelled — no GPU instance was created.\"; exit 1; }\n\nbrev create digital-health-clinical-asr-sft \\\n  --gpu l40s:1 --image ubuntu-22-04-cuda-12-4 --disk 200gi\nbrev ssh-config                             # writes ~\u002F.ssh\u002Fconfig entries\nrsync -avz .\u002Fcycle1\u002F digital-health-clinical-asr-sft:~\u002Fcycle1\u002F\nbrev shell digital-health-clinical-asr-sft            # drops into the instance\nnvidia-smi                                  # confirm GPU\ndocker pull nvcr.io\u002Fnvidia\u002Fnemo:25.11.01    # ~12 GB, once per instance\n```\n\nWhen 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\u002Fcontainer-paths.md`.\n\n### 4b. Term-aware train\u002Fval split\n\n**Row-disjoint, stratified by `entity_category`, default val fraction 0.2.**\n\nThe **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.\n\nSingleton categories (one row total) get forced to train with a warning. If any priority category has \u003C 5 rows, **bail to `\u002Fdigital-health-clinical-asr-build`** — held-out validation will be too noisy to attribute movement.\n\nSketch:\n\n```python\n# After loading manifest.jsonl into a list of dicts `rows`:\nfrom collections import defaultdict\nimport random\nrandom.seed(42)\n\nby_cat = defaultdict(list)\nfor r in rows:\n    by_cat[r[\"entity_category\"]].append(r)\n\ntrain, val = [], []\nfor cat, cat_rows in by_cat.items():\n    random.shuffle(cat_rows)\n    if len(cat_rows) \u003C 2:\n        train.extend(cat_rows)\n        print(f\"warning: singleton category {cat}, forced to train\")\n        continue\n    n_val = max(1, int(0.2 * len(cat_rows)))\n    val.extend(cat_rows[:n_val])\n    train.extend(cat_rows[n_val:])\n```\n\nWrite `train.jsonl` and `validation.jsonl` alongside the manifest. **These are the inputs to `speech_to_text_finetune.py`.**\n\n### 4c. Choose the base model\n\n| Base | SFT viability | Notes |\n|---|---|---|\n| **`nvidia\u002Fparakeet-tdt-0.6b-v2`** | ✅ **Empirically verified** (KER 0.513 → 0.128 in 3 epochs, −75% relative) | NVIDIA's current English ASR default. Stock NeMo SFT recipe works end-to-end. **Recommended.** |\n| `nvidia\u002Fnemotron-speech-streaming-en-0.6b` | ❌ **Don't use for SFT** | NVCF function is streaming-only; SFT path unreliable (UNK collapse on validation after first training step). For streaming serving, Riva chunks a non-streaming base just fine. |\n\nOther Parakeet\u002FConformer bases (1.1B, CTC, RNNT, `stt_en_conformer_ctc_large`) + decoder → NIM container mapping: `references\u002Fstage4-finetune.md`. If the user asks to fine-tune Nemotron Speech Streaming, **warn about the collapse and recommend Parakeet TDT v2**.\n\n### 4d. Stock NeMo SFT\n\nIn the NeMo container, invoke `\u002Fopt\u002FNeMo\u002Fexamples\u002Fasr\u002Fspeech_to_text_finetune.py` directly. **No custom adapter logic. No patches.** The stock NeMo SFT script is the verified working recipe.\n\nHyperparameters (verified on Parakeet TDT v2, 39-row manifest):\n\n```\ninit_from_pretrained_model: nvidia\u002Fparakeet-tdt-0.6b-v2\nprecision:                  bf16-mixed       # required for TDT numerical stability\nlr:                         3e-4             # CosineAnnealing schedule\nwarmup_steps:               5                # tiny manifest; bump to 500 at production scale\nepochs:                     3                # smoke; 10-30 for production\nbatch_size:                 4                # fits 16 GB VRAM; raise to 16 on L40S 48 GB\ngradient_clip_val:          1.0              # defensive\n```\n\n**Container invocation**: `docker run --gpus all --rm -it -v \"$PWD:\u002Fworkspace\" nvcr.io\u002Fnvidia\u002Fnemo:25.11.01 python \u002Fopt\u002FNeMo\u002Fexamples\u002Fasr\u002Fspeech_to_text_finetune.py` with `model.train_ds.manifest_filepath=\u002Fworkspace\u002Ftrain.jsonl`, `model.validation_ds.manifest_filepath=\u002Fworkspace\u002Fvalidation.jsonl`, `init_from_pretrained_model=nvidia\u002Fparakeet-tdt-0.6b-v2`, and the hyperparameter overrides from the table above. Full docker-run line with config-path \u002F config-name flags: `references\u002Fstage4-finetune.md` §Container invocation.\n\n**Manifest paths inside the container.** Host paths (e.g. `$HOME\u002F…`) don't resolve in `\u002Fworkspace`. Rewrite snippet: `references\u002Fcontainer-paths.md`.\n\nThe 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\u003CN>\u002Fmodels\u002F\u003Crun>\u002F`; the layout doesn't matter as long as it's consistent across cycles).\n\n### 4e. Offline cycle N+1 eval — close the loop\n\nRe-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.\n\nSketch:\n\n```python\nimport nemo.collections.asr as nemo_asr\nmodel = nemo_asr.models.ASRModel.restore_from(\"adapted_model.nemo\")\nhyps = model.transcribe([\"audio\u002Frow1.wav\", \"audio\u002Frow2.wav\", ...])\n```\n\nScore the same four metrics (WER\u002FCER\u002FKER\u002FSER) and the same five-section leaderboard the eval skill produces. Write them as `leaderboard_cycle\u003CN+1>.md`. Compare against `leaderboard_cycle\u003CN>.md`.\n\n**Decision table** — cycle-N+1 vs cycle-N:\n\n| Result | Action |\n|---|---|\n| KER dropped meaningfully on targeted categories (e.g. drug KER −20% or more, relative) | ✅ Keep the `.nemo`. Update the leaderboard. Advance to Step 4f if you want to deploy. |\n| KER moved a little, you wanted more | Loop back to `\u002Fdigital-health-clinical-asr-build`, expand the manifest. Tiny manifests rarely benefit from hyperparameter tweaks — signal density beats LR sweeps. |\n| KER got worse | Overfit on a tiny manifest. Bail to `\u002Fdigital-health-clinical-asr-build` and grow before retraining. Don't tune harder on the same data. |\n| No measurable change | Some categories may already be in the base model's vocab. Sanity-check per-category numbers before concluding training \"didn't help.\" |\n\n### 4f. (Optional) Deploy as a Riva NIM\n\nHand the `.nemo` to `\u002Friva-asr-custom`. **Pass the source architecture explicitly** — `\u002Friva-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:\n\n| Source decoder | `riva-build` flag | NIM container family |\n|---|---|---|\n| Conformer-CTC | `decoder=greedy_ctc` | `parakeet-*-ctc-*` |\n| Conformer-RNNT | `decoder=nemo` | `parakeet-rnnt-*` |\n| **Conformer-TDT (default)** | `decoder=nemo` | `parakeet-tdt-*` |\n| Cache-Aware RNNT (Nemotron streaming) | `decoder=nemo` | `nemotron-streaming-*` ⚠ SFT broken on this base, see Limitations |\n\nAfter deploy: re-run `\u002Fdigital-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 `\u002Friva-asr-custom`.\n\n## Examples\n\n**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 `\u002Fdigital-health-clinical-asr-build`.\n\n**Scenario B — Nemotron Streaming.** User: *\"SFT `nvidia\u002Fnemotron-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.\n\n**Scenario C — cycle 2 KER unchanged.** User: *\"KER barely moved.\"* → Back to `\u002Fdigital-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 — `\u002Fdigital-health-clinical-asr-build` Step 2d.\n\n## Artifacts produced\n\n- `train.jsonl`, `validation.jsonl` — term-aware split (Step 4b)\n- `adapted_model.nemo` — fine-tuned model (Step 4d)\n- `training_run_info.json` — hyperparameters, dataset stats, end-of-train metrics\n- `offline_hyps.jsonl` — cycle-N+1 transcription hypotheses (Step 4e)\n- `leaderboard_cycle\u003CN+1>.md` — cycle-N+1 five-section leaderboard\n- *(optional, after Step 4f)* a deployed NIM endpoint (delegated to `\u002Friva-asr-custom`)\n\n## Troubleshooting\n\n- **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\u002Fparakeet-tdt-0.6b-v2` (the recommended default) or `nvidia\u002Fstt_en_conformer_ctc_large` (legacy fallback). The streaming RNNT SFT path is broken; do not retry with different hyperparameters.\n- **Manifest paths don't resolve inside the NeMo container** → host paths (e.g. `$HOME\u002F…`) need rewriting to `\u002Fworkspace\u002F…`. See `references\u002Fcontainer-paths.md` for the rewrite snippet.\n- **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.)\n- **Cycle N+1 KER got worse** → overfit on a tiny manifest. Bail to `\u002Fdigital-health-clinical-asr-build` and grow.\n- **Riva-served numbers diverge from offline numbers** → the gap is in Riva preprocessing or `riva-build` flags, not the model. Route to `\u002Friva-asr-custom`.\n- **`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.\n- **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.\n\n## Limitations\n\n- **Adapter-style SFT on TDT\u002FRNNT 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\u002FRNNT bases.\n- **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.\n- **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.\n- **English-only by default.** The base-model table is en-US-specific. Other locales need a different base + a re-validated SFT recipe.\n- **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.\n\n## Next steps\n\n- **Deploy the `.nemo` as a NIM:** `\u002Friva-asr-custom` (pass the source architecture explicitly).\n- **Grow the manifest for cycle N+2:** `\u002Fdigital-health-clinical-asr-build`.\n- **Re-score the cycle:** `\u002Fdigital-health-clinical-asr-eval` (against the new endpoint or the new `.nemo` directly).\n- **Lateral** for word boosting \u002F LM fusion \u002F non-clinical SFT recipes: `\u002Ffinetune-asr`.\n\n## References\n\n- [`references\u002Fstage4-finetune.md`](references\u002Fstage4-finetune.md) — base-model selection table, hyperparameter rationale, decoder → NIM container mapping, decision tree comparing cycle-N+1 to cycle-N\n- [`references\u002Fcontainer-paths.md`](references\u002Fcontainer-paths.md) — host → `\u002Fworkspace\u002F` path rewriting for cross-host manifest portability (laptop ↔ Brev ↔ NeMo container)\n\n\n",{"data":38,"body":62},{"name":4,"description":6,"version":39,"author":40,"tags":41,"tools":48,"license":29,"compatibility":53,"metadata":54},"1.0.0","Ben Randoing \u003Cbrandoing@nvidia.com>",[42,43,44,45,46,47],"clinical-asr","finetune","sft","nemo","parakeet","flywheel",[49,50,51,52],"Read","Write","Bash","Skill","Requires a CUDA host (24 GB VRAM comfortable, 16 GB workable with batch_size=4), the NeMo container (nvcr.io\u002Fnvidia\u002Fnemo:25.11.01), and the finetune-asr + riva-asr-custom skills installed alongside this one. No local GPU? Use Brev. NVIDIA_API_KEY required for the offline cycle N+1 eval round-trip and for any NIM deploy.",{"author":40,"tags":55,"team":57,"domain":58,"stage":59,"previous_skill":60,"next_skill":61},[42,47,43,56,46],"nemo-sft","healthcare-tme","ai-ml",4,"digital-health-clinical-asr-eval","riva-asr-custom",{"type":63,"children":64},"root",[65,74,90,119,162,172,199,206,211,328,334,373,379,384,443,455,506,512,638,644,651,663,726,738,765,1191,1225,1231,1246,1264,1281,1286,1449,1482,1488,1583,1609,1615,1634,1639,1649,1695,1727,1756,1762,1789,1793,1824,1844,1854,1949,1955,1993,2133,2166,2172,2210,2240,2286,2292,2369,2375,2577,2583,2656,2662,2743,2749,2787],{"type":66,"tag":67,"props":68,"children":70},"element","h1",{"id":69},"clinical-asr-flywheel-stage-4-fine-tune",[71],{"type":72,"value":73},"text","Clinical ASR Flywheel — Stage 4 (Fine-tune)",{"type":66,"tag":75,"props":76,"children":77},"blockquote",{},[78],{"type":66,"tag":79,"props":80,"children":81},"p",{},[82,88],{"type":66,"tag":83,"props":84,"children":85},"strong",{},[86],{"type":72,"value":87},"⚠ Agent: read this entire SKILL.md before answering.",{"type":72,"value":89}," 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.",{"type":66,"tag":75,"props":91,"children":92},{},[93],{"type":66,"tag":79,"props":94,"children":95},{},[96,101,103,108,110,117],{"type":66,"tag":83,"props":97,"children":98},{},[99],{"type":72,"value":100},"Agent: this file is self-contained.",{"type":72,"value":102}," The Stage 4 gate criteria, base-model recommendation, hyperparameter table, container invocation pattern, and cycle-N+1 decision table are all below. ",{"type":66,"tag":83,"props":104,"children":105},{},[106],{"type":72,"value":107},"Do not",{"type":72,"value":109}," run file-discovery commands or open ",{"type":66,"tag":111,"props":112,"children":114},"code",{"className":113},[],[115],{"type":72,"value":116},"references\u002Fstage4-finetune.md",{"type":72,"value":118}," 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.",{"type":66,"tag":79,"props":120,"children":121},{},[122,124,129,131,137,139,144,146,152,154,160],{"type":72,"value":123},"You are the ",{"type":66,"tag":83,"props":125,"children":126},{},[127],{"type":72,"value":128},"adapt-and-measure",{"type":72,"value":130}," stage. The user arrives from ",{"type":66,"tag":111,"props":132,"children":134},{"className":133},[],[135],{"type":72,"value":136},"\u002Fdigital-health-clinical-asr-eval",{"type":72,"value":138}," 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 ",{"type":66,"tag":83,"props":140,"children":141},{},[142],{"type":72,"value":143},"measure that the loop closed",{"type":72,"value":145},", and optionally hand the resulting ",{"type":66,"tag":111,"props":147,"children":149},{"className":148},[],[150],{"type":72,"value":151},".nemo",{"type":72,"value":153}," to ",{"type":66,"tag":111,"props":155,"children":157},{"className":156},[],[158],{"type":72,"value":159},"\u002Friva-asr-custom",{"type":72,"value":161}," for production serving.",{"type":66,"tag":79,"props":163,"children":164},{},[165,170],{"type":66,"tag":83,"props":166,"children":167},{},[168],{"type":72,"value":169},"The cycle KER from offline eval is the measurement that closes the loop.",{"type":72,"value":171}," Riva NIM deploy validates serving (latency, streaming, scale), not model quality.",{"type":66,"tag":75,"props":173,"children":174},{},[175],{"type":66,"tag":79,"props":176,"children":177},{},[178,183,185,190,192,197],{"type":66,"tag":83,"props":179,"children":180},{},[181],{"type":72,"value":182},"Empirically verified on the reference manifest",{"type":72,"value":184}," (39 rows, Parakeet TDT v2):\nBaseline KER ",{"type":66,"tag":83,"props":186,"children":187},{},[188],{"type":72,"value":189},"0.513",{"type":72,"value":191}," → after 3 epochs of stock SFT: ",{"type":66,"tag":83,"props":193,"children":194},{},[195],{"type":72,"value":196},"0.128",{"type":72,"value":198}," (-75% relative).\nDrug names: 0.857 → 0.214. Conditions: 0.500 → 0.000. Procedures: 0.250 → 0.000.",{"type":66,"tag":200,"props":201,"children":203},"h2",{"id":202},"critical-workflow-rules-apply-on-every-activation",[204],{"type":72,"value":205},"Critical workflow rules (apply on every activation)",{"type":66,"tag":79,"props":207,"children":208},{},[209],{"type":72,"value":210},"Surface these facts in any response, even if the user asks a narrow question:",{"type":66,"tag":212,"props":213,"children":214},"ol",{},[215,226,243,269,286,303],{"type":66,"tag":216,"props":217,"children":218},"li",{},[219,224],{"type":66,"tag":83,"props":220,"children":221},{},[222],{"type":72,"value":223},"Read this entire SKILL.md before answering.",{"type":72,"value":225}," The base-model selection table, hyperparameter values, and the cycle-N+1 decision table are below — they are the load-bearing parts.",{"type":66,"tag":216,"props":227,"children":228},{},[229,234,236,241],{"type":66,"tag":83,"props":230,"children":231},{},[232],{"type":72,"value":233},"Verified result",{"type":72,"value":235}," — Parakeet TDT v2 with the recipe in §4c achieves ",{"type":66,"tag":83,"props":237,"children":238},{},[239],{"type":72,"value":240},"KER 0.513 → 0.128 (−75% relative)",{"type":72,"value":242}," in 3 epochs on the reference manifest. Cite this when the user asks whether SFT will help.",{"type":66,"tag":216,"props":244,"children":245},{},[246,267],{"type":66,"tag":83,"props":247,"children":248},{},[249,251,257,259,265],{"type":72,"value":250},"Recipe is ",{"type":66,"tag":111,"props":252,"children":254},{"className":253},[],[255],{"type":72,"value":256},"\u002Fopt\u002FNeMo\u002Fexamples\u002Fasr\u002Fspeech_to_text_finetune.py",{"type":72,"value":258}," inside ",{"type":66,"tag":111,"props":260,"children":262},{"className":261},[],[263],{"type":72,"value":264},"nvcr.io\u002Fnvidia\u002Fnemo:25.11.01",{"type":72,"value":266},".",{"type":72,"value":268}," Stock script, no patches, no custom adapter logic. The adapter-mixin path is broken on TDT\u002FRNNT decoders (72 NaN tensors at any LR) — do not propose it.",{"type":66,"tag":216,"props":270,"children":271},{},[272,284],{"type":66,"tag":83,"props":273,"children":274},{},[275,277,283],{"type":72,"value":276},"Recommended base is ",{"type":66,"tag":111,"props":278,"children":280},{"className":279},[],[281],{"type":72,"value":282},"nvidia\u002Fparakeet-tdt-0.6b-v2",{"type":72,"value":266},{"type":72,"value":285}," The full base-model table is in §4c.",{"type":66,"tag":216,"props":287,"children":288},{},[289,301],{"type":66,"tag":83,"props":290,"children":291},{},[292,294,300],{"type":72,"value":293},"Do NOT fine-tune ",{"type":66,"tag":111,"props":295,"children":297},{"className":296},[],[298],{"type":72,"value":299},"nvidia\u002Fnemotron-speech-streaming-en-0.6b",{"type":72,"value":266},{"type":72,"value":302}," 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.",{"type":66,"tag":216,"props":304,"children":305},{},[306,311,313,318,320,326],{"type":66,"tag":83,"props":307,"children":308},{},[309],{"type":72,"value":310},"Gate the recommendation.",{"type":72,"value":312}," Stage 4 only fires when priority-category KER > 0.3 ",{"type":66,"tag":83,"props":314,"children":315},{},[316],{"type":72,"value":317},"and",{"type":72,"value":319}," manifest has ≥ 100 rows (≥ 5 per priority category). Below those thresholds, route back to ",{"type":66,"tag":111,"props":321,"children":323},{"className":322},[],[324],{"type":72,"value":325},"\u002Fdigital-health-clinical-asr-build",{"type":72,"value":327}," to grow the manifest first.",{"type":66,"tag":200,"props":329,"children":331},{"id":330},"purpose",[332],{"type":72,"value":333},"Purpose",{"type":66,"tag":79,"props":335,"children":336},{},[337,339,344,346,351,353,358,360,365,366,371],{"type":72,"value":338},"Run ",{"type":66,"tag":83,"props":340,"children":341},{},[342],{"type":72,"value":343},"stock NeMo SFT",{"type":72,"value":345}," (no custom adapter logic, no patches) in ",{"type":66,"tag":111,"props":347,"children":349},{"className":348},[],[350],{"type":72,"value":264},{"type":72,"value":352}," against a term-aware row-disjoint train\u002Fval split, produce a ",{"type":66,"tag":111,"props":354,"children":356},{"className":355},[],[357],{"type":72,"value":151},{"type":72,"value":359}," 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 ",{"type":66,"tag":111,"props":361,"children":363},{"className":362},[],[364],{"type":72,"value":151},{"type":72,"value":153},{"type":66,"tag":111,"props":367,"children":369},{"className":368},[],[370],{"type":72,"value":159},{"type":72,"value":372}," for NIM deploy.",{"type":66,"tag":200,"props":374,"children":376},{"id":375},"when-to-use-this-skill",[377],{"type":72,"value":378},"When to use this skill",{"type":66,"tag":79,"props":380,"children":381},{},[382],{"type":72,"value":383},"Activate on user phrases like:",{"type":66,"tag":385,"props":386,"children":387},"ul",{},[388,393,398,403,408,413,418],{"type":66,"tag":216,"props":389,"children":390},{},[391],{"type":72,"value":392},"\"Fine-tune ASR on my clinical vocabulary\"",{"type":66,"tag":216,"props":394,"children":395},{},[396],{"type":72,"value":397},"\"Improve ASR on medication names\"",{"type":66,"tag":216,"props":399,"children":400},{},[401],{"type":72,"value":402},"\"We have a KER of 0.4, can we fine-tune?\"",{"type":66,"tag":216,"props":404,"children":405},{},[406],{"type":72,"value":407},"\"Run SFT on my Parakeet TDT base\"",{"type":66,"tag":216,"props":409,"children":410},{},[411],{"type":72,"value":412},"\"Train a clinical ASR adapter\"",{"type":66,"tag":216,"props":414,"children":415},{},[416],{"type":72,"value":417},"\"Compare cycle 1 vs cycle 2 KER\"",{"type":66,"tag":216,"props":419,"children":420},{},[421,423],{"type":72,"value":422},"\"Deploy my fine-tuned model as a NIM\" ",{"type":66,"tag":424,"props":425,"children":426},"em",{},[427,429,434,436,441],{"type":72,"value":428},"(this skill prepares the ",{"type":66,"tag":111,"props":430,"children":432},{"className":431},[],[433],{"type":72,"value":151},{"type":72,"value":435}," and routes to ",{"type":66,"tag":111,"props":437,"children":439},{"className":438},[],[440],{"type":72,"value":159},{"type":72,"value":442}," for the deploy)",{"type":66,"tag":79,"props":444,"children":445},{},[446,448,453],{"type":72,"value":447},"Do ",{"type":66,"tag":83,"props":449,"children":450},{},[451],{"type":72,"value":452},"not",{"type":72,"value":454}," activate when:",{"type":66,"tag":385,"props":456,"children":457},{},[458,468,478,489],{"type":66,"tag":216,"props":459,"children":460},{},[461,463],{"type":72,"value":462},"The user hasn't scored a baseline yet → ",{"type":66,"tag":111,"props":464,"children":466},{"className":465},[],[467],{"type":72,"value":136},{"type":66,"tag":216,"props":469,"children":470},{},[471,473],{"type":72,"value":472},"The user doesn't have a manifest → ",{"type":66,"tag":111,"props":474,"children":476},{"className":475},[],[477],{"type":72,"value":325},{"type":66,"tag":216,"props":479,"children":480},{},[481,483],{"type":72,"value":482},"The user wants generic word boosting \u002F LM fusion (not SFT) → ",{"type":66,"tag":111,"props":484,"children":486},{"className":485},[],[487],{"type":72,"value":488},"\u002Ffinetune-asr",{"type":66,"tag":216,"props":490,"children":491},{},[492,494,499,501],{"type":72,"value":493},"The user has a ",{"type":66,"tag":111,"props":495,"children":497},{"className":496},[],[498],{"type":72,"value":151},{"type":72,"value":500}," and only wants to deploy → ",{"type":66,"tag":111,"props":502,"children":504},{"className":503},[],[505],{"type":72,"value":159},{"type":66,"tag":200,"props":507,"children":509},{"id":508},"prerequisites",[510],{"type":72,"value":511},"Prerequisites",{"type":66,"tag":385,"props":513,"children":514},{},[515,540,566,590,608,625],{"type":66,"tag":216,"props":516,"children":517},{},[518,523,525,530,532,538],{"type":66,"tag":83,"props":519,"children":520},{},[521],{"type":72,"value":522},"A cycle-N manifest + cycle-N eval result",{"type":72,"value":524}," from ",{"type":66,"tag":111,"props":526,"children":528},{"className":527},[],[529],{"type":72,"value":136},{"type":72,"value":531},". The priority-category KER must be > 0.3 (Stage 4 gate). The manifest should have ≥ 100 rows total, and ≥ 5 rows per priority ",{"type":66,"tag":111,"props":533,"children":535},{"className":534},[],[536],{"type":72,"value":537},"entity_category",{"type":72,"value":539},", for a believable post-tune signal.",{"type":66,"tag":216,"props":541,"children":542},{},[543,548,550,556,558,564],{"type":66,"tag":83,"props":544,"children":545},{},[546],{"type":72,"value":547},"A CUDA host",{"type":72,"value":549}," — 24 GB VRAM is comfortable for Parakeet TDT 0.6B at ",{"type":66,"tag":111,"props":551,"children":553},{"className":552},[],[554],{"type":72,"value":555},"batch_size=4",{"type":72,"value":557}," with ",{"type":66,"tag":111,"props":559,"children":561},{"className":560},[],[562],{"type":72,"value":563},"bf16-mixed",{"type":72,"value":565},"; 16 GB works with smaller batch. No local GPU? Use Brev — recommended SKU is L40S 48 GB.",{"type":66,"tag":216,"props":567,"children":568},{},[569,574,576,581,583,589],{"type":66,"tag":83,"props":570,"children":571},{},[572],{"type":72,"value":573},"The NeMo container",{"type":72,"value":575},": ",{"type":66,"tag":111,"props":577,"children":579},{"className":578},[],[580],{"type":72,"value":264},{"type":72,"value":582},". Pull once: ",{"type":66,"tag":111,"props":584,"children":586},{"className":585},[],[587],{"type":72,"value":588},"docker pull nvcr.io\u002Fnvidia\u002Fnemo:25.11.01",{"type":72,"value":266},{"type":66,"tag":216,"props":591,"children":592},{},[593,598,600,606],{"type":66,"tag":83,"props":594,"children":595},{},[596],{"type":72,"value":597},"NVIDIA Container Toolkit + Docker",{"type":72,"value":599}," — covered by ",{"type":66,"tag":111,"props":601,"children":603},{"className":602},[],[604],{"type":72,"value":605},"\u002Friva-nim-setup",{"type":72,"value":607}," if not already installed.",{"type":66,"tag":216,"props":609,"children":610},{},[611,616,618,623],{"type":66,"tag":83,"props":612,"children":613},{},[614],{"type":72,"value":615},"A train\u002Fval split",{"type":72,"value":617}," stratified by ",{"type":66,"tag":111,"props":619,"children":621},{"className":620},[],[622],{"type":72,"value":537},{"type":72,"value":624}," (recipe sketch in Step 4b below).",{"type":66,"tag":216,"props":626,"children":627},{},[628,636],{"type":66,"tag":83,"props":629,"children":630},{},[631],{"type":66,"tag":111,"props":632,"children":634},{"className":633},[],[635],{"type":72,"value":159},{"type":72,"value":637}," installed if you intend to deploy. Pure-research SFT runs without it.",{"type":66,"tag":200,"props":639,"children":641},{"id":640},"instructions",[642],{"type":72,"value":643},"Instructions",{"type":66,"tag":645,"props":646,"children":648},"h3",{"id":647},"_4a-provision-a-gpu-host-skip-if-you-already-have-one",[649],{"type":72,"value":650},"4a. Provision a GPU host (skip if you already have one)",{"type":66,"tag":79,"props":652,"children":653},{},[654,656,661],{"type":72,"value":655},"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 ",{"type":66,"tag":83,"props":657,"children":658},{},[659],{"type":72,"value":660},"Brev",{"type":72,"value":662}," — NVIDIA's per-second-billed GPU host service. Recommended SKU: L40S 48 GB.",{"type":66,"tag":79,"props":664,"children":665},{},[666,678,680,686,688,693,695,701,703,709,711,716,718,724],{"type":66,"tag":83,"props":667,"children":668},{},[669,671,677],{"type":72,"value":670},"Cost disclosure — surface this to the user before any ",{"type":66,"tag":111,"props":672,"children":674},{"className":673},[],[675],{"type":72,"value":676},"brev create",{"type":72,"value":266},{"type":72,"value":679}," L40S 48 GB runs ",{"type":66,"tag":681,"props":682,"children":683},"del",{},[684],{"type":72,"value":685},"$1.50\u002Fhr at time of writing; a 3-epoch SFT run on a 100-row manifest finishes in 15–30 minutes (",{"type":72,"value":687},"$0.40–$0.75 of compute). The real risk is ",{"type":66,"tag":83,"props":689,"children":690},{},[691],{"type":72,"value":692},"forgetting to stop the instance",{"type":72,"value":694}," — overnight idle on L40S is ~$36, a week of idle is ~$250. Mitigations: (a) always wrap the workflow in a script that ends with ",{"type":66,"tag":111,"props":696,"children":698},{"className":697},[],[699],{"type":72,"value":700},"brev stop",{"type":72,"value":702},"; (b) set a calendar reminder when you start; (c) ",{"type":66,"tag":111,"props":704,"children":706},{"className":705},[],[707],{"type":72,"value":708},"brev delete",{"type":72,"value":710}," instead of ",{"type":66,"tag":111,"props":712,"children":714},{"className":713},[],[715],{"type":72,"value":700},{"type":72,"value":717}," if you don't need to keep the disk (",{"type":66,"tag":111,"props":719,"children":721},{"className":720},[],[722],{"type":72,"value":723},"stop",{"type":72,"value":725}," keeps disk at $0.10\u002FGB-month — 200 GB ≈ $20\u002Fmonth of latent cost). Confirm the user accepts the per-hour cost shape and the idle risk before spinning anything up.",{"type":66,"tag":79,"props":727,"children":728},{},[729,731,736],{"type":72,"value":730},"Full setup walkthrough — CLI install (download-then-run, not curl-pipe), SKU choice, disk sizing, SSH config — is in ",{"type":66,"tag":111,"props":732,"children":734},{"className":733},[],[735],{"type":72,"value":116},{"type":72,"value":737}," (§Brev provisioning).",{"type":66,"tag":79,"props":739,"children":740},{},[741,743,763],{"type":72,"value":742},"Short happy-path once the CLI is installed. ",{"type":66,"tag":83,"props":744,"children":745},{},[746,748,753,755,761],{"type":72,"value":747},"Do not run ",{"type":66,"tag":111,"props":749,"children":751},{"className":750},[],[752],{"type":72,"value":676},{"type":72,"value":754}," until the user has explicitly typed ",{"type":66,"tag":111,"props":756,"children":758},{"className":757},[],[759],{"type":72,"value":760},"YES",{"type":72,"value":762}," at the confirmation prompt below",{"type":72,"value":764}," — the gate is mandatory, not advisory, because everything after it bills against the user's account by the second:",{"type":66,"tag":766,"props":767,"children":772},"pre",{"className":768,"code":769,"language":770,"meta":771,"style":771},"language-bash shiki shiki-themes material-theme-lighter material-theme material-theme-palenight","brev login                                  # browser auth\n\n# Mandatory cost-confirmation gate — do NOT skip or auto-answer this.\necho \"About to provision: digital-health-clinical-asr-sft on L40S 48 GB.\"\necho \"Cost shape: ~\\$1.50\u002Fhr while running; ~\\$36\u002Fnight if left idle; ~\\$20\u002Fmo disk if you 'stop' instead of 'delete'.\"\nread -rp \"Type YES to provision (anything else cancels): \" confirm\n[ \"$confirm\" = \"YES\" ] || { echo \"Cancelled — no GPU instance was created.\"; exit 1; }\n\nbrev create digital-health-clinical-asr-sft \\\n  --gpu l40s:1 --image ubuntu-22-04-cuda-12-4 --disk 200gi\nbrev ssh-config                             # writes ~\u002F.ssh\u002Fconfig entries\nrsync -avz .\u002Fcycle1\u002F digital-health-clinical-asr-sft:~\u002Fcycle1\u002F\nbrev shell digital-health-clinical-asr-sft            # drops into the instance\nnvidia-smi                                  # confirm GPU\ndocker pull nvcr.io\u002Fnvidia\u002Fnemo:25.11.01    # ~12 GB, once per instance\n","bash","",[773],{"type":66,"tag":111,"props":774,"children":775},{"__ignoreMap":771},[776,800,810,819,844,894,927,1024,1032,1055,1089,1107,1131,1153,1167],{"type":66,"tag":777,"props":778,"children":781},"span",{"class":779,"line":780},"line",1,[782,788,794],{"type":66,"tag":777,"props":783,"children":785},{"style":784},"--shiki-light:#E2931D;--shiki-default:#FFCB6B;--shiki-dark:#FFCB6B",[786],{"type":72,"value":787},"brev",{"type":66,"tag":777,"props":789,"children":791},{"style":790},"--shiki-light:#91B859;--shiki-default:#C3E88D;--shiki-dark:#C3E88D",[792],{"type":72,"value":793}," login",{"type":66,"tag":777,"props":795,"children":797},{"style":796},"--shiki-light:#90A4AE;--shiki-light-font-style:italic;--shiki-default:#546E7A;--shiki-default-font-style:italic;--shiki-dark:#676E95;--shiki-dark-font-style:italic",[798],{"type":72,"value":799},"                                  # browser auth\n",{"type":66,"tag":777,"props":801,"children":803},{"class":779,"line":802},2,[804],{"type":66,"tag":777,"props":805,"children":807},{"emptyLinePlaceholder":806},true,[808],{"type":72,"value":809},"\n",{"type":66,"tag":777,"props":811,"children":813},{"class":779,"line":812},3,[814],{"type":66,"tag":777,"props":815,"children":816},{"style":796},[817],{"type":72,"value":818},"# Mandatory cost-confirmation gate — do NOT skip or auto-answer this.\n",{"type":66,"tag":777,"props":820,"children":821},{"class":779,"line":59},[822,828,834,839],{"type":66,"tag":777,"props":823,"children":825},{"style":824},"--shiki-light:#6182B8;--shiki-default:#82AAFF;--shiki-dark:#82AAFF",[826],{"type":72,"value":827},"echo",{"type":66,"tag":777,"props":829,"children":831},{"style":830},"--shiki-light:#39ADB5;--shiki-default:#89DDFF;--shiki-dark:#89DDFF",[832],{"type":72,"value":833}," \"",{"type":66,"tag":777,"props":835,"children":836},{"style":790},[837],{"type":72,"value":838},"About to provision: digital-health-clinical-asr-sft on L40S 48 GB.",{"type":66,"tag":777,"props":840,"children":841},{"style":830},[842],{"type":72,"value":843},"\"\n",{"type":66,"tag":777,"props":845,"children":847},{"class":779,"line":846},5,[848,852,856,861,867,872,876,881,885,890],{"type":66,"tag":777,"props":849,"children":850},{"style":824},[851],{"type":72,"value":827},{"type":66,"tag":777,"props":853,"children":854},{"style":830},[855],{"type":72,"value":833},{"type":66,"tag":777,"props":857,"children":858},{"style":790},[859],{"type":72,"value":860},"Cost shape: ~",{"type":66,"tag":777,"props":862,"children":864},{"style":863},"--shiki-light:#90A4AE;--shiki-default:#EEFFFF;--shiki-dark:#BABED8",[865],{"type":72,"value":866},"\\$",{"type":66,"tag":777,"props":868,"children":869},{"style":790},[870],{"type":72,"value":871},"1.50\u002Fhr while running; ~",{"type":66,"tag":777,"props":873,"children":874},{"style":863},[875],{"type":72,"value":866},{"type":66,"tag":777,"props":877,"children":878},{"style":790},[879],{"type":72,"value":880},"36\u002Fnight if left idle; ~",{"type":66,"tag":777,"props":882,"children":883},{"style":863},[884],{"type":72,"value":866},{"type":66,"tag":777,"props":886,"children":887},{"style":790},[888],{"type":72,"value":889},"20\u002Fmo disk if you 'stop' instead of 'delete'.",{"type":66,"tag":777,"props":891,"children":892},{"style":830},[893],{"type":72,"value":843},{"type":66,"tag":777,"props":895,"children":897},{"class":779,"line":896},6,[898,903,908,912,917,922],{"type":66,"tag":777,"props":899,"children":900},{"style":824},[901],{"type":72,"value":902},"read",{"type":66,"tag":777,"props":904,"children":905},{"style":790},[906],{"type":72,"value":907}," -rp",{"type":66,"tag":777,"props":909,"children":910},{"style":830},[911],{"type":72,"value":833},{"type":66,"tag":777,"props":913,"children":914},{"style":790},[915],{"type":72,"value":916},"Type YES to provision (anything else cancels): ",{"type":66,"tag":777,"props":918,"children":919},{"style":830},[920],{"type":72,"value":921},"\"",{"type":66,"tag":777,"props":923,"children":924},{"style":790},[925],{"type":72,"value":926}," confirm\n",{"type":66,"tag":777,"props":928,"children":930},{"class":779,"line":929},7,[931,936,940,945,949,954,958,962,966,971,976,981,986,990,995,999,1004,1009,1015,1019],{"type":66,"tag":777,"props":932,"children":933},{"style":830},[934],{"type":72,"value":935},"[",{"type":66,"tag":777,"props":937,"children":938},{"style":830},[939],{"type":72,"value":833},{"type":66,"tag":777,"props":941,"children":942},{"style":863},[943],{"type":72,"value":944},"$confirm",{"type":66,"tag":777,"props":946,"children":947},{"style":830},[948],{"type":72,"value":921},{"type":66,"tag":777,"props":950,"children":951},{"style":830},[952],{"type":72,"value":953}," =",{"type":66,"tag":777,"props":955,"children":956},{"style":830},[957],{"type":72,"value":833},{"type":66,"tag":777,"props":959,"children":960},{"style":790},[961],{"type":72,"value":760},{"type":66,"tag":777,"props":963,"children":964},{"style":830},[965],{"type":72,"value":921},{"type":66,"tag":777,"props":967,"children":968},{"style":830},[969],{"type":72,"value":970}," ]",{"type":66,"tag":777,"props":972,"children":973},{"style":830},[974],{"type":72,"value":975}," ||",{"type":66,"tag":777,"props":977,"children":978},{"style":830},[979],{"type":72,"value":980}," {",{"type":66,"tag":777,"props":982,"children":983},{"style":824},[984],{"type":72,"value":985}," echo",{"type":66,"tag":777,"props":987,"children":988},{"style":830},[989],{"type":72,"value":833},{"type":66,"tag":777,"props":991,"children":992},{"style":790},[993],{"type":72,"value":994},"Cancelled — no GPU instance was created.",{"type":66,"tag":777,"props":996,"children":997},{"style":830},[998],{"type":72,"value":921},{"type":66,"tag":777,"props":1000,"children":1001},{"style":830},[1002],{"type":72,"value":1003},";",{"type":66,"tag":777,"props":1005,"children":1006},{"style":824},[1007],{"type":72,"value":1008}," exit",{"type":66,"tag":777,"props":1010,"children":1012},{"style":1011},"--shiki-light:#F76D47;--shiki-default:#F78C6C;--shiki-dark:#F78C6C",[1013],{"type":72,"value":1014}," 1",{"type":66,"tag":777,"props":1016,"children":1017},{"style":830},[1018],{"type":72,"value":1003},{"type":66,"tag":777,"props":1020,"children":1021},{"style":830},[1022],{"type":72,"value":1023}," }\n",{"type":66,"tag":777,"props":1025,"children":1027},{"class":779,"line":1026},8,[1028],{"type":66,"tag":777,"props":1029,"children":1030},{"emptyLinePlaceholder":806},[1031],{"type":72,"value":809},{"type":66,"tag":777,"props":1033,"children":1035},{"class":779,"line":1034},9,[1036,1040,1045,1050],{"type":66,"tag":777,"props":1037,"children":1038},{"style":784},[1039],{"type":72,"value":787},{"type":66,"tag":777,"props":1041,"children":1042},{"style":790},[1043],{"type":72,"value":1044}," create",{"type":66,"tag":777,"props":1046,"children":1047},{"style":790},[1048],{"type":72,"value":1049}," digital-health-clinical-asr-sft",{"type":66,"tag":777,"props":1051,"children":1052},{"style":863},[1053],{"type":72,"value":1054}," \\\n",{"type":66,"tag":777,"props":1056,"children":1058},{"class":779,"line":1057},10,[1059,1064,1069,1074,1079,1084],{"type":66,"tag":777,"props":1060,"children":1061},{"style":790},[1062],{"type":72,"value":1063},"  --gpu",{"type":66,"tag":777,"props":1065,"children":1066},{"style":790},[1067],{"type":72,"value":1068}," l40s:1",{"type":66,"tag":777,"props":1070,"children":1071},{"style":790},[1072],{"type":72,"value":1073}," --image",{"type":66,"tag":777,"props":1075,"children":1076},{"style":790},[1077],{"type":72,"value":1078}," ubuntu-22-04-cuda-12-4",{"type":66,"tag":777,"props":1080,"children":1081},{"style":790},[1082],{"type":72,"value":1083}," --disk",{"type":66,"tag":777,"props":1085,"children":1086},{"style":790},[1087],{"type":72,"value":1088}," 200gi\n",{"type":66,"tag":777,"props":1090,"children":1092},{"class":779,"line":1091},11,[1093,1097,1102],{"type":66,"tag":777,"props":1094,"children":1095},{"style":784},[1096],{"type":72,"value":787},{"type":66,"tag":777,"props":1098,"children":1099},{"style":790},[1100],{"type":72,"value":1101}," ssh-config",{"type":66,"tag":777,"props":1103,"children":1104},{"style":796},[1105],{"type":72,"value":1106},"                             # writes ~\u002F.ssh\u002Fconfig entries\n",{"type":66,"tag":777,"props":1108,"children":1110},{"class":779,"line":1109},12,[1111,1116,1121,1126],{"type":66,"tag":777,"props":1112,"children":1113},{"style":784},[1114],{"type":72,"value":1115},"rsync",{"type":66,"tag":777,"props":1117,"children":1118},{"style":790},[1119],{"type":72,"value":1120}," -avz",{"type":66,"tag":777,"props":1122,"children":1123},{"style":790},[1124],{"type":72,"value":1125}," .\u002Fcycle1\u002F",{"type":66,"tag":777,"props":1127,"children":1128},{"style":790},[1129],{"type":72,"value":1130}," digital-health-clinical-asr-sft:~\u002Fcycle1\u002F\n",{"type":66,"tag":777,"props":1132,"children":1134},{"class":779,"line":1133},13,[1135,1139,1144,1148],{"type":66,"tag":777,"props":1136,"children":1137},{"style":784},[1138],{"type":72,"value":787},{"type":66,"tag":777,"props":1140,"children":1141},{"style":790},[1142],{"type":72,"value":1143}," shell",{"type":66,"tag":777,"props":1145,"children":1146},{"style":790},[1147],{"type":72,"value":1049},{"type":66,"tag":777,"props":1149,"children":1150},{"style":796},[1151],{"type":72,"value":1152},"            # drops into the instance\n",{"type":66,"tag":777,"props":1154,"children":1156},{"class":779,"line":1155},14,[1157,1162],{"type":66,"tag":777,"props":1158,"children":1159},{"style":784},[1160],{"type":72,"value":1161},"nvidia-smi",{"type":66,"tag":777,"props":1163,"children":1164},{"style":796},[1165],{"type":72,"value":1166},"                                  # confirm GPU\n",{"type":66,"tag":777,"props":1168,"children":1170},{"class":779,"line":1169},15,[1171,1176,1181,1186],{"type":66,"tag":777,"props":1172,"children":1173},{"style":784},[1174],{"type":72,"value":1175},"docker",{"type":66,"tag":777,"props":1177,"children":1178},{"style":790},[1179],{"type":72,"value":1180}," pull",{"type":66,"tag":777,"props":1182,"children":1183},{"style":790},[1184],{"type":72,"value":1185}," nvcr.io\u002Fnvidia\u002Fnemo:25.11.01",{"type":66,"tag":777,"props":1187,"children":1188},{"style":796},[1189],{"type":72,"value":1190},"    # ~12 GB, once per instance\n",{"type":66,"tag":79,"props":1192,"children":1193},{},[1194,1196,1201,1202,1208,1210,1216,1218,1224],{"type":72,"value":1195},"When done, ",{"type":66,"tag":83,"props":1197,"children":1198},{},[1199],{"type":72,"value":1200},"always halt billing",{"type":72,"value":575},{"type":66,"tag":111,"props":1203,"children":1205},{"className":1204},[],[1206],{"type":72,"value":1207},"brev stop digital-health-clinical-asr-sft",{"type":72,"value":1209}," (keeps disk) or ",{"type":66,"tag":111,"props":1211,"children":1213},{"className":1212},[],[1214],{"type":72,"value":1215},"brev delete digital-health-clinical-asr-sft",{"type":72,"value":1217}," (frees it). For path rewriting laptop → Brev → NeMo container, see ",{"type":66,"tag":111,"props":1219,"children":1221},{"className":1220},[],[1222],{"type":72,"value":1223},"references\u002Fcontainer-paths.md",{"type":72,"value":266},{"type":66,"tag":645,"props":1226,"children":1228},{"id":1227},"_4b-term-aware-trainval-split",[1229],{"type":72,"value":1230},"4b. Term-aware train\u002Fval split",{"type":66,"tag":79,"props":1232,"children":1233},{},[1234],{"type":66,"tag":83,"props":1235,"children":1236},{},[1237,1239,1244],{"type":72,"value":1238},"Row-disjoint, stratified by ",{"type":66,"tag":111,"props":1240,"children":1242},{"className":1241},[],[1243],{"type":72,"value":537},{"type":72,"value":1245},", default val fraction 0.2.",{"type":66,"tag":79,"props":1247,"children":1248},{},[1249,1251,1262],{"type":72,"value":1250},"The ",{"type":66,"tag":83,"props":1252,"children":1253},{},[1254,1256],{"type":72,"value":1255},"same ",{"type":66,"tag":111,"props":1257,"children":1259},{"className":1258},[],[1260],{"type":72,"value":1261},"term",{"type":72,"value":1263}," 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.",{"type":66,"tag":79,"props":1265,"children":1266},{},[1267,1269,1279],{"type":72,"value":1268},"Singleton categories (one row total) get forced to train with a warning. If any priority category has \u003C 5 rows, ",{"type":66,"tag":83,"props":1270,"children":1271},{},[1272,1274],{"type":72,"value":1273},"bail to ",{"type":66,"tag":111,"props":1275,"children":1277},{"className":1276},[],[1278],{"type":72,"value":325},{"type":72,"value":1280}," — held-out validation will be too noisy to attribute movement.",{"type":66,"tag":79,"props":1282,"children":1283},{},[1284],{"type":72,"value":1285},"Sketch:",{"type":66,"tag":766,"props":1287,"children":1291},{"className":1288,"code":1289,"language":1290,"meta":771,"style":771},"language-python shiki shiki-themes material-theme-lighter material-theme material-theme-palenight","# After loading manifest.jsonl into a list of dicts `rows`:\nfrom collections import defaultdict\nimport random\nrandom.seed(42)\n\nby_cat = defaultdict(list)\nfor r in rows:\n    by_cat[r[\"entity_category\"]].append(r)\n\ntrain, val = [], []\nfor cat, cat_rows in by_cat.items():\n    random.shuffle(cat_rows)\n    if len(cat_rows) \u003C 2:\n        train.extend(cat_rows)\n        print(f\"warning: singleton category {cat}, forced to train\")\n        continue\n    n_val = max(1, int(0.2 * len(cat_rows)))\n    val.extend(cat_rows[:n_val])\n    train.extend(cat_rows[n_val:])\n","python",[1292],{"type":66,"tag":111,"props":1293,"children":1294},{"__ignoreMap":771},[1295,1303,1311,1319,1327,1334,1342,1350,1358,1365,1373,1381,1389,1397,1405,1413,1422,1431,1440],{"type":66,"tag":777,"props":1296,"children":1297},{"class":779,"line":780},[1298],{"type":66,"tag":777,"props":1299,"children":1300},{},[1301],{"type":72,"value":1302},"# After loading manifest.jsonl into a list of dicts `rows`:\n",{"type":66,"tag":777,"props":1304,"children":1305},{"class":779,"line":802},[1306],{"type":66,"tag":777,"props":1307,"children":1308},{},[1309],{"type":72,"value":1310},"from collections import defaultdict\n",{"type":66,"tag":777,"props":1312,"children":1313},{"class":779,"line":812},[1314],{"type":66,"tag":777,"props":1315,"children":1316},{},[1317],{"type":72,"value":1318},"import random\n",{"type":66,"tag":777,"props":1320,"children":1321},{"class":779,"line":59},[1322],{"type":66,"tag":777,"props":1323,"children":1324},{},[1325],{"type":72,"value":1326},"random.seed(42)\n",{"type":66,"tag":777,"props":1328,"children":1329},{"class":779,"line":846},[1330],{"type":66,"tag":777,"props":1331,"children":1332},{"emptyLinePlaceholder":806},[1333],{"type":72,"value":809},{"type":66,"tag":777,"props":1335,"children":1336},{"class":779,"line":896},[1337],{"type":66,"tag":777,"props":1338,"children":1339},{},[1340],{"type":72,"value":1341},"by_cat = defaultdict(list)\n",{"type":66,"tag":777,"props":1343,"children":1344},{"class":779,"line":929},[1345],{"type":66,"tag":777,"props":1346,"children":1347},{},[1348],{"type":72,"value":1349},"for r in rows:\n",{"type":66,"tag":777,"props":1351,"children":1352},{"class":779,"line":1026},[1353],{"type":66,"tag":777,"props":1354,"children":1355},{},[1356],{"type":72,"value":1357},"    by_cat[r[\"entity_category\"]].append(r)\n",{"type":66,"tag":777,"props":1359,"children":1360},{"class":779,"line":1034},[1361],{"type":66,"tag":777,"props":1362,"children":1363},{"emptyLinePlaceholder":806},[1364],{"type":72,"value":809},{"type":66,"tag":777,"props":1366,"children":1367},{"class":779,"line":1057},[1368],{"type":66,"tag":777,"props":1369,"children":1370},{},[1371],{"type":72,"value":1372},"train, val = [], []\n",{"type":66,"tag":777,"props":1374,"children":1375},{"class":779,"line":1091},[1376],{"type":66,"tag":777,"props":1377,"children":1378},{},[1379],{"type":72,"value":1380},"for cat, cat_rows in by_cat.items():\n",{"type":66,"tag":777,"props":1382,"children":1383},{"class":779,"line":1109},[1384],{"type":66,"tag":777,"props":1385,"children":1386},{},[1387],{"type":72,"value":1388},"    random.shuffle(cat_rows)\n",{"type":66,"tag":777,"props":1390,"children":1391},{"class":779,"line":1133},[1392],{"type":66,"tag":777,"props":1393,"children":1394},{},[1395],{"type":72,"value":1396},"    if len(cat_rows) \u003C 2:\n",{"type":66,"tag":777,"props":1398,"children":1399},{"class":779,"line":1155},[1400],{"type":66,"tag":777,"props":1401,"children":1402},{},[1403],{"type":72,"value":1404},"        train.extend(cat_rows)\n",{"type":66,"tag":777,"props":1406,"children":1407},{"class":779,"line":1169},[1408],{"type":66,"tag":777,"props":1409,"children":1410},{},[1411],{"type":72,"value":1412},"        print(f\"warning: singleton category {cat}, forced to train\")\n",{"type":66,"tag":777,"props":1414,"children":1416},{"class":779,"line":1415},16,[1417],{"type":66,"tag":777,"props":1418,"children":1419},{},[1420],{"type":72,"value":1421},"        continue\n",{"type":66,"tag":777,"props":1423,"children":1425},{"class":779,"line":1424},17,[1426],{"type":66,"tag":777,"props":1427,"children":1428},{},[1429],{"type":72,"value":1430},"    n_val = max(1, int(0.2 * len(cat_rows)))\n",{"type":66,"tag":777,"props":1432,"children":1434},{"class":779,"line":1433},18,[1435],{"type":66,"tag":777,"props":1436,"children":1437},{},[1438],{"type":72,"value":1439},"    val.extend(cat_rows[:n_val])\n",{"type":66,"tag":777,"props":1441,"children":1443},{"class":779,"line":1442},19,[1444],{"type":66,"tag":777,"props":1445,"children":1446},{},[1447],{"type":72,"value":1448},"    train.extend(cat_rows[n_val:])\n",{"type":66,"tag":79,"props":1450,"children":1451},{},[1452,1454,1460,1462,1468,1470],{"type":72,"value":1453},"Write ",{"type":66,"tag":111,"props":1455,"children":1457},{"className":1456},[],[1458],{"type":72,"value":1459},"train.jsonl",{"type":72,"value":1461}," and ",{"type":66,"tag":111,"props":1463,"children":1465},{"className":1464},[],[1466],{"type":72,"value":1467},"validation.jsonl",{"type":72,"value":1469}," alongside the manifest. ",{"type":66,"tag":83,"props":1471,"children":1472},{},[1473,1475,1481],{"type":72,"value":1474},"These are the inputs to ",{"type":66,"tag":111,"props":1476,"children":1478},{"className":1477},[],[1479],{"type":72,"value":1480},"speech_to_text_finetune.py",{"type":72,"value":266},{"type":66,"tag":645,"props":1483,"children":1485},{"id":1484},"_4c-choose-the-base-model",[1486],{"type":72,"value":1487},"4c. Choose the base model",{"type":66,"tag":1489,"props":1490,"children":1491},"table",{},[1492,1516],{"type":66,"tag":1493,"props":1494,"children":1495},"thead",{},[1496],{"type":66,"tag":1497,"props":1498,"children":1499},"tr",{},[1500,1506,1511],{"type":66,"tag":1501,"props":1502,"children":1503},"th",{},[1504],{"type":72,"value":1505},"Base",{"type":66,"tag":1501,"props":1507,"children":1508},{},[1509],{"type":72,"value":1510},"SFT viability",{"type":66,"tag":1501,"props":1512,"children":1513},{},[1514],{"type":72,"value":1515},"Notes",{"type":66,"tag":1517,"props":1518,"children":1519},"tbody",{},[1520,1557],{"type":66,"tag":1497,"props":1521,"children":1522},{},[1523,1535,1547],{"type":66,"tag":1524,"props":1525,"children":1526},"td",{},[1527],{"type":66,"tag":83,"props":1528,"children":1529},{},[1530],{"type":66,"tag":111,"props":1531,"children":1533},{"className":1532},[],[1534],{"type":72,"value":282},{"type":66,"tag":1524,"props":1536,"children":1537},{},[1538,1540,1545],{"type":72,"value":1539},"✅ ",{"type":66,"tag":83,"props":1541,"children":1542},{},[1543],{"type":72,"value":1544},"Empirically verified",{"type":72,"value":1546}," (KER 0.513 → 0.128 in 3 epochs, −75% relative)",{"type":66,"tag":1524,"props":1548,"children":1549},{},[1550,1552],{"type":72,"value":1551},"NVIDIA's current English ASR default. Stock NeMo SFT recipe works end-to-end. ",{"type":66,"tag":83,"props":1553,"children":1554},{},[1555],{"type":72,"value":1556},"Recommended.",{"type":66,"tag":1497,"props":1558,"children":1559},{},[1560,1568,1578],{"type":66,"tag":1524,"props":1561,"children":1562},{},[1563],{"type":66,"tag":111,"props":1564,"children":1566},{"className":1565},[],[1567],{"type":72,"value":299},{"type":66,"tag":1524,"props":1569,"children":1570},{},[1571,1573],{"type":72,"value":1572},"❌ ",{"type":66,"tag":83,"props":1574,"children":1575},{},[1576],{"type":72,"value":1577},"Don't use for SFT",{"type":66,"tag":1524,"props":1579,"children":1580},{},[1581],{"type":72,"value":1582},"NVCF function is streaming-only; SFT path unreliable (UNK collapse on validation after first training step). For streaming serving, Riva chunks a non-streaming base just fine.",{"type":66,"tag":79,"props":1584,"children":1585},{},[1586,1588,1594,1596,1601,1603,1608],{"type":72,"value":1587},"Other Parakeet\u002FConformer bases (1.1B, CTC, RNNT, ",{"type":66,"tag":111,"props":1589,"children":1591},{"className":1590},[],[1592],{"type":72,"value":1593},"stt_en_conformer_ctc_large",{"type":72,"value":1595},") + decoder → NIM container mapping: ",{"type":66,"tag":111,"props":1597,"children":1599},{"className":1598},[],[1600],{"type":72,"value":116},{"type":72,"value":1602},". If the user asks to fine-tune Nemotron Speech Streaming, ",{"type":66,"tag":83,"props":1604,"children":1605},{},[1606],{"type":72,"value":1607},"warn about the collapse and recommend Parakeet TDT v2",{"type":72,"value":266},{"type":66,"tag":645,"props":1610,"children":1612},{"id":1611},"_4d-stock-nemo-sft",[1613],{"type":72,"value":1614},"4d. Stock NeMo SFT",{"type":66,"tag":79,"props":1616,"children":1617},{},[1618,1620,1625,1627,1632],{"type":72,"value":1619},"In the NeMo container, invoke ",{"type":66,"tag":111,"props":1621,"children":1623},{"className":1622},[],[1624],{"type":72,"value":256},{"type":72,"value":1626}," directly. ",{"type":66,"tag":83,"props":1628,"children":1629},{},[1630],{"type":72,"value":1631},"No custom adapter logic. No patches.",{"type":72,"value":1633}," The stock NeMo SFT script is the verified working recipe.",{"type":66,"tag":79,"props":1635,"children":1636},{},[1637],{"type":72,"value":1638},"Hyperparameters (verified on Parakeet TDT v2, 39-row manifest):",{"type":66,"tag":766,"props":1640,"children":1644},{"className":1641,"code":1643,"language":72},[1642],"language-text","init_from_pretrained_model: nvidia\u002Fparakeet-tdt-0.6b-v2\nprecision:                  bf16-mixed       # required for TDT numerical stability\nlr:                         3e-4             # CosineAnnealing schedule\nwarmup_steps:               5                # tiny manifest; bump to 500 at production scale\nepochs:                     3                # smoke; 10-30 for production\nbatch_size:                 4                # fits 16 GB VRAM; raise to 16 on L40S 48 GB\ngradient_clip_val:          1.0              # defensive\n",[1645],{"type":66,"tag":111,"props":1646,"children":1647},{"__ignoreMap":771},[1648],{"type":72,"value":1643},{"type":66,"tag":79,"props":1650,"children":1651},{},[1652,1657,1658,1664,1665,1671,1673,1679,1680,1686,1688,1693],{"type":66,"tag":83,"props":1653,"children":1654},{},[1655],{"type":72,"value":1656},"Container invocation",{"type":72,"value":575},{"type":66,"tag":111,"props":1659,"children":1661},{"className":1660},[],[1662],{"type":72,"value":1663},"docker run --gpus all --rm -it -v \"$PWD:\u002Fworkspace\" nvcr.io\u002Fnvidia\u002Fnemo:25.11.01 python \u002Fopt\u002FNeMo\u002Fexamples\u002Fasr\u002Fspeech_to_text_finetune.py",{"type":72,"value":557},{"type":66,"tag":111,"props":1666,"children":1668},{"className":1667},[],[1669],{"type":72,"value":1670},"model.train_ds.manifest_filepath=\u002Fworkspace\u002Ftrain.jsonl",{"type":72,"value":1672},", ",{"type":66,"tag":111,"props":1674,"children":1676},{"className":1675},[],[1677],{"type":72,"value":1678},"model.validation_ds.manifest_filepath=\u002Fworkspace\u002Fvalidation.jsonl",{"type":72,"value":1672},{"type":66,"tag":111,"props":1681,"children":1683},{"className":1682},[],[1684],{"type":72,"value":1685},"init_from_pretrained_model=nvidia\u002Fparakeet-tdt-0.6b-v2",{"type":72,"value":1687},", and the hyperparameter overrides from the table above. Full docker-run line with config-path \u002F config-name flags: ",{"type":66,"tag":111,"props":1689,"children":1691},{"className":1690},[],[1692],{"type":72,"value":116},{"type":72,"value":1694}," §Container invocation.",{"type":66,"tag":79,"props":1696,"children":1697},{},[1698,1703,1705,1711,1713,1719,1721,1726],{"type":66,"tag":83,"props":1699,"children":1700},{},[1701],{"type":72,"value":1702},"Manifest paths inside the container.",{"type":72,"value":1704}," Host paths (e.g. ",{"type":66,"tag":111,"props":1706,"children":1708},{"className":1707},[],[1709],{"type":72,"value":1710},"$HOME\u002F…",{"type":72,"value":1712},") don't resolve in ",{"type":66,"tag":111,"props":1714,"children":1716},{"className":1715},[],[1717],{"type":72,"value":1718},"\u002Fworkspace",{"type":72,"value":1720},". Rewrite snippet: ",{"type":66,"tag":111,"props":1722,"children":1724},{"className":1723},[],[1725],{"type":72,"value":1223},{"type":72,"value":266},{"type":66,"tag":79,"props":1728,"children":1729},{},[1730,1732,1738,1740,1746,1748,1754],{"type":72,"value":1731},"The training run writes ",{"type":66,"tag":111,"props":1733,"children":1735},{"className":1734},[],[1736],{"type":72,"value":1737},"adapted_model.nemo",{"type":72,"value":1739}," and a ",{"type":66,"tag":111,"props":1741,"children":1743},{"className":1742},[],[1744],{"type":72,"value":1745},"training_run_info.json",{"type":72,"value":1747}," summary. Both go into a per-cycle subdirectory of the user's choice (e.g. ",{"type":66,"tag":111,"props":1749,"children":1751},{"className":1750},[],[1752],{"type":72,"value":1753},"cycle\u003CN>\u002Fmodels\u002F\u003Crun>\u002F",{"type":72,"value":1755},"; the layout doesn't matter as long as it's consistent across cycles).",{"type":66,"tag":645,"props":1757,"children":1759},{"id":1758},"_4e-offline-cycle-n1-eval-close-the-loop",[1760],{"type":72,"value":1761},"4e. Offline cycle N+1 eval — close the loop",{"type":66,"tag":79,"props":1763,"children":1764},{},[1765,1767,1772,1774,1780,1782,1787],{"type":72,"value":1766},"Re-transcribe the cycle's audio with the fine-tuned ",{"type":66,"tag":111,"props":1768,"children":1770},{"className":1769},[],[1771],{"type":72,"value":151},{"type":72,"value":1773}," using NeMo's offline ",{"type":66,"tag":111,"props":1775,"children":1777},{"className":1776},[],[1778],{"type":72,"value":1779},"transcribe()",{"type":72,"value":1781},". ",{"type":66,"tag":83,"props":1783,"children":1784},{},[1785],{"type":72,"value":1786},"No Riva needed",{"type":72,"value":1788}," — this is measurement, not serving. NeMo's offline path runs the same encoder + decoder graph the Riva NIM eventually serves.",{"type":66,"tag":79,"props":1790,"children":1791},{},[1792],{"type":72,"value":1285},{"type":66,"tag":766,"props":1794,"children":1796},{"className":1288,"code":1795,"language":1290,"meta":771,"style":771},"import nemo.collections.asr as nemo_asr\nmodel = nemo_asr.models.ASRModel.restore_from(\"adapted_model.nemo\")\nhyps = model.transcribe([\"audio\u002Frow1.wav\", \"audio\u002Frow2.wav\", ...])\n",[1797],{"type":66,"tag":111,"props":1798,"children":1799},{"__ignoreMap":771},[1800,1808,1816],{"type":66,"tag":777,"props":1801,"children":1802},{"class":779,"line":780},[1803],{"type":66,"tag":777,"props":1804,"children":1805},{},[1806],{"type":72,"value":1807},"import nemo.collections.asr as nemo_asr\n",{"type":66,"tag":777,"props":1809,"children":1810},{"class":779,"line":802},[1811],{"type":66,"tag":777,"props":1812,"children":1813},{},[1814],{"type":72,"value":1815},"model = nemo_asr.models.ASRModel.restore_from(\"adapted_model.nemo\")\n",{"type":66,"tag":777,"props":1817,"children":1818},{"class":779,"line":812},[1819],{"type":66,"tag":777,"props":1820,"children":1821},{},[1822],{"type":72,"value":1823},"hyps = model.transcribe([\"audio\u002Frow1.wav\", \"audio\u002Frow2.wav\", ...])\n",{"type":66,"tag":79,"props":1825,"children":1826},{},[1827,1829,1835,1837,1843],{"type":72,"value":1828},"Score the same four metrics (WER\u002FCER\u002FKER\u002FSER) and the same five-section leaderboard the eval skill produces. Write them as ",{"type":66,"tag":111,"props":1830,"children":1832},{"className":1831},[],[1833],{"type":72,"value":1834},"leaderboard_cycle\u003CN+1>.md",{"type":72,"value":1836},". Compare against ",{"type":66,"tag":111,"props":1838,"children":1840},{"className":1839},[],[1841],{"type":72,"value":1842},"leaderboard_cycle\u003CN>.md",{"type":72,"value":266},{"type":66,"tag":79,"props":1845,"children":1846},{},[1847,1852],{"type":66,"tag":83,"props":1848,"children":1849},{},[1850],{"type":72,"value":1851},"Decision table",{"type":72,"value":1853}," — cycle-N+1 vs cycle-N:",{"type":66,"tag":1489,"props":1855,"children":1856},{},[1857,1873],{"type":66,"tag":1493,"props":1858,"children":1859},{},[1860],{"type":66,"tag":1497,"props":1861,"children":1862},{},[1863,1868],{"type":66,"tag":1501,"props":1864,"children":1865},{},[1866],{"type":72,"value":1867},"Result",{"type":66,"tag":1501,"props":1869,"children":1870},{},[1871],{"type":72,"value":1872},"Action",{"type":66,"tag":1517,"props":1874,"children":1875},{},[1876,1896,1916,1936],{"type":66,"tag":1497,"props":1877,"children":1878},{},[1879,1884],{"type":66,"tag":1524,"props":1880,"children":1881},{},[1882],{"type":72,"value":1883},"KER dropped meaningfully on targeted categories (e.g. drug KER −20% or more, relative)",{"type":66,"tag":1524,"props":1885,"children":1886},{},[1887,1889,1894],{"type":72,"value":1888},"✅ Keep the ",{"type":66,"tag":111,"props":1890,"children":1892},{"className":1891},[],[1893],{"type":72,"value":151},{"type":72,"value":1895},". Update the leaderboard. Advance to Step 4f if you want to deploy.",{"type":66,"tag":1497,"props":1897,"children":1898},{},[1899,1904],{"type":66,"tag":1524,"props":1900,"children":1901},{},[1902],{"type":72,"value":1903},"KER moved a little, you wanted more",{"type":66,"tag":1524,"props":1905,"children":1906},{},[1907,1909,1914],{"type":72,"value":1908},"Loop back to ",{"type":66,"tag":111,"props":1910,"children":1912},{"className":1911},[],[1913],{"type":72,"value":325},{"type":72,"value":1915},", expand the manifest. Tiny manifests rarely benefit from hyperparameter tweaks — signal density beats LR sweeps.",{"type":66,"tag":1497,"props":1917,"children":1918},{},[1919,1924],{"type":66,"tag":1524,"props":1920,"children":1921},{},[1922],{"type":72,"value":1923},"KER got worse",{"type":66,"tag":1524,"props":1925,"children":1926},{},[1927,1929,1934],{"type":72,"value":1928},"Overfit on a tiny manifest. Bail to ",{"type":66,"tag":111,"props":1930,"children":1932},{"className":1931},[],[1933],{"type":72,"value":325},{"type":72,"value":1935}," and grow before retraining. Don't tune harder on the same data.",{"type":66,"tag":1497,"props":1937,"children":1938},{},[1939,1944],{"type":66,"tag":1524,"props":1940,"children":1941},{},[1942],{"type":72,"value":1943},"No measurable change",{"type":66,"tag":1524,"props":1945,"children":1946},{},[1947],{"type":72,"value":1948},"Some categories may already be in the base model's vocab. Sanity-check per-category numbers before concluding training \"didn't help.\"",{"type":66,"tag":645,"props":1950,"children":1952},{"id":1951},"_4f-optional-deploy-as-a-riva-nim",[1953],{"type":72,"value":1954},"4f. (Optional) Deploy as a Riva NIM",{"type":66,"tag":79,"props":1956,"children":1957},{},[1958,1960,1965,1966,1971,1972,1977,1979,1984,1986,1991],{"type":72,"value":1959},"Hand the ",{"type":66,"tag":111,"props":1961,"children":1963},{"className":1962},[],[1964],{"type":72,"value":151},{"type":72,"value":153},{"type":66,"tag":111,"props":1967,"children":1969},{"className":1968},[],[1970],{"type":72,"value":159},{"type":72,"value":1781},{"type":66,"tag":83,"props":1973,"children":1974},{},[1975],{"type":72,"value":1976},"Pass the source architecture explicitly",{"type":72,"value":1978}," — ",{"type":66,"tag":111,"props":1980,"children":1982},{"className":1981},[],[1983],{"type":72,"value":159},{"type":72,"value":1985}," can't reliably detect CTC vs RNNT vs TDT from the ",{"type":66,"tag":111,"props":1987,"children":1989},{"className":1988},[],[1990],{"type":72,"value":151},{"type":72,"value":1992}," alone, and the wrong NIM container produces a broken RMIR with no clear error:",{"type":66,"tag":1489,"props":1994,"children":1995},{},[1996,2023],{"type":66,"tag":1493,"props":1997,"children":1998},{},[1999],{"type":66,"tag":1497,"props":2000,"children":2001},{},[2002,2007,2018],{"type":66,"tag":1501,"props":2003,"children":2004},{},[2005],{"type":72,"value":2006},"Source decoder",{"type":66,"tag":1501,"props":2008,"children":2009},{},[2010,2016],{"type":66,"tag":111,"props":2011,"children":2013},{"className":2012},[],[2014],{"type":72,"value":2015},"riva-build",{"type":72,"value":2017}," flag",{"type":66,"tag":1501,"props":2019,"children":2020},{},[2021],{"type":72,"value":2022},"NIM container family",{"type":66,"tag":1517,"props":2024,"children":2025},{},[2026,2052,2078,2106],{"type":66,"tag":1497,"props":2027,"children":2028},{},[2029,2034,2043],{"type":66,"tag":1524,"props":2030,"children":2031},{},[2032],{"type":72,"value":2033},"Conformer-CTC",{"type":66,"tag":1524,"props":2035,"children":2036},{},[2037],{"type":66,"tag":111,"props":2038,"children":2040},{"className":2039},[],[2041],{"type":72,"value":2042},"decoder=greedy_ctc",{"type":66,"tag":1524,"props":2044,"children":2045},{},[2046],{"type":66,"tag":111,"props":2047,"children":2049},{"className":2048},[],[2050],{"type":72,"value":2051},"parakeet-*-ctc-*",{"type":66,"tag":1497,"props":2053,"children":2054},{},[2055,2060,2069],{"type":66,"tag":1524,"props":2056,"children":2057},{},[2058],{"type":72,"value":2059},"Conformer-RNNT",{"type":66,"tag":1524,"props":2061,"children":2062},{},[2063],{"type":66,"tag":111,"props":2064,"children":2066},{"className":2065},[],[2067],{"type":72,"value":2068},"decoder=nemo",{"type":66,"tag":1524,"props":2070,"children":2071},{},[2072],{"type":66,"tag":111,"props":2073,"children":2075},{"className":2074},[],[2076],{"type":72,"value":2077},"parakeet-rnnt-*",{"type":66,"tag":1497,"props":2079,"children":2080},{},[2081,2089,2097],{"type":66,"tag":1524,"props":2082,"children":2083},{},[2084],{"type":66,"tag":83,"props":2085,"children":2086},{},[2087],{"type":72,"value":2088},"Conformer-TDT (default)",{"type":66,"tag":1524,"props":2090,"children":2091},{},[2092],{"type":66,"tag":111,"props":2093,"children":2095},{"className":2094},[],[2096],{"type":72,"value":2068},{"type":66,"tag":1524,"props":2098,"children":2099},{},[2100],{"type":66,"tag":111,"props":2101,"children":2103},{"className":2102},[],[2104],{"type":72,"value":2105},"parakeet-tdt-*",{"type":66,"tag":1497,"props":2107,"children":2108},{},[2109,2114,2122],{"type":66,"tag":1524,"props":2110,"children":2111},{},[2112],{"type":72,"value":2113},"Cache-Aware RNNT (Nemotron streaming)",{"type":66,"tag":1524,"props":2115,"children":2116},{},[2117],{"type":66,"tag":111,"props":2118,"children":2120},{"className":2119},[],[2121],{"type":72,"value":2068},{"type":66,"tag":1524,"props":2123,"children":2124},{},[2125,2131],{"type":66,"tag":111,"props":2126,"children":2128},{"className":2127},[],[2129],{"type":72,"value":2130},"nemotron-streaming-*",{"type":72,"value":2132}," ⚠ SFT broken on this base, see Limitations",{"type":66,"tag":79,"props":2134,"children":2135},{},[2136,2138,2143,2145,2151,2153,2158,2160,2165],{"type":72,"value":2137},"After deploy: re-run ",{"type":66,"tag":111,"props":2139,"children":2141},{"className":2140},[],[2142],{"type":72,"value":136},{"type":72,"value":2144}," against the new endpoint (",{"type":66,"tag":111,"props":2146,"children":2148},{"className":2147},[],[2149],{"type":72,"value":2150},"ASR_ENDPOINT=localhost:50051",{"type":72,"value":2152},") to validate that production-serving numbers match offline numbers. Any divergence is in Riva preprocessing or ",{"type":66,"tag":111,"props":2154,"children":2156},{"className":2155},[],[2157],{"type":72,"value":2015},{"type":72,"value":2159}," flags, not the model. Route to ",{"type":66,"tag":111,"props":2161,"children":2163},{"className":2162},[],[2164],{"type":72,"value":159},{"type":72,"value":266},{"type":66,"tag":200,"props":2167,"children":2169},{"id":2168},"examples",[2170],{"type":72,"value":2171},"Examples",{"type":66,"tag":79,"props":2173,"children":2174},{},[2175,2180,2182,2187,2189,2195,2197,2202,2204,2209],{"type":66,"tag":83,"props":2176,"children":2177},{},[2178],{"type":72,"value":2179},"Scenario A — gate met.",{"type":72,"value":2181}," User: ",{"type":66,"tag":424,"props":2183,"children":2184},{},[2185],{"type":72,"value":2186},"\"Drug KER 0.42, 130 rows. SFT?\"",{"type":72,"value":2188}," → Yes (gate cleared). ",{"type":66,"tag":111,"props":2190,"children":2192},{"className":2191},[],[2193],{"type":72,"value":2194},"parakeet-tdt-0.6b-v2",{"type":72,"value":2196}," (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 ",{"type":66,"tag":111,"props":2198,"children":2200},{"className":2199},[],[2201],{"type":72,"value":151},{"type":72,"value":2203},"; otherwise back to ",{"type":66,"tag":111,"props":2205,"children":2207},{"className":2206},[],[2208],{"type":72,"value":325},{"type":72,"value":266},{"type":66,"tag":79,"props":2211,"children":2212},{},[2213,2218,2219,2231,2233,2238],{"type":66,"tag":83,"props":2214,"children":2215},{},[2216],{"type":72,"value":2217},"Scenario B — Nemotron Streaming.",{"type":72,"value":2181},{"type":66,"tag":424,"props":2220,"children":2221},{},[2222,2224,2229],{"type":72,"value":2223},"\"SFT ",{"type":66,"tag":111,"props":2225,"children":2227},{"className":2226},[],[2228],{"type":72,"value":299},{"type":72,"value":2230},"?\"",{"type":72,"value":2232}," → No (UNK collapse). Substitute ",{"type":66,"tag":111,"props":2234,"children":2236},{"className":2235},[],[2237],{"type":72,"value":2194},{"type":72,"value":2239},". Riva chunks non-streaming bases for streaming serving — base doesn't need to be streaming-native.",{"type":66,"tag":79,"props":2241,"children":2242},{},[2243,2248,2249,2254,2256,2261,2263,2269,2271,2277,2279,2284],{"type":66,"tag":83,"props":2244,"children":2245},{},[2246],{"type":72,"value":2247},"Scenario C — cycle 2 KER unchanged.",{"type":72,"value":2181},{"type":66,"tag":424,"props":2250,"children":2251},{},[2252],{"type":72,"value":2253},"\"KER barely moved.\"",{"type":72,"value":2255}," → Back to ",{"type":66,"tag":111,"props":2257,"children":2259},{"className":2258},[],[2260],{"type":72,"value":325},{"type":72,"value":2262},". Signal density beats LR sweeps. If ",{"type":66,"tag":111,"props":2264,"children":2266},{"className":2265},[],[2267],{"type":72,"value":2268},"magpie_g2p",{"type":72,"value":2270}," rows are bad but ",{"type":66,"tag":111,"props":2272,"children":2274},{"className":2273},[],[2275],{"type":72,"value":2276},"merriam-webster",{"type":72,"value":2278}," rows are good, the gap is pronunciation-coverage — ",{"type":66,"tag":111,"props":2280,"children":2282},{"className":2281},[],[2283],{"type":72,"value":325},{"type":72,"value":2285}," Step 2d.",{"type":66,"tag":200,"props":2287,"children":2289},{"id":2288},"artifacts-produced",[2290],{"type":72,"value":2291},"Artifacts produced",{"type":66,"tag":385,"props":2293,"children":2294},{},[2295,2311,2321,2331,2342,2352],{"type":66,"tag":216,"props":2296,"children":2297},{},[2298,2303,2304,2309],{"type":66,"tag":111,"props":2299,"children":2301},{"className":2300},[],[2302],{"type":72,"value":1459},{"type":72,"value":1672},{"type":66,"tag":111,"props":2305,"children":2307},{"className":2306},[],[2308],{"type":72,"value":1467},{"type":72,"value":2310}," — term-aware split (Step 4b)",{"type":66,"tag":216,"props":2312,"children":2313},{},[2314,2319],{"type":66,"tag":111,"props":2315,"children":2317},{"className":2316},[],[2318],{"type":72,"value":1737},{"type":72,"value":2320}," — fine-tuned model (Step 4d)",{"type":66,"tag":216,"props":2322,"children":2323},{},[2324,2329],{"type":66,"tag":111,"props":2325,"children":2327},{"className":2326},[],[2328],{"type":72,"value":1745},{"type":72,"value":2330}," — hyperparameters, dataset stats, end-of-train metrics",{"type":66,"tag":216,"props":2332,"children":2333},{},[2334,2340],{"type":66,"tag":111,"props":2335,"children":2337},{"className":2336},[],[2338],{"type":72,"value":2339},"offline_hyps.jsonl",{"type":72,"value":2341}," — cycle-N+1 transcription hypotheses (Step 4e)",{"type":66,"tag":216,"props":2343,"children":2344},{},[2345,2350],{"type":66,"tag":111,"props":2346,"children":2348},{"className":2347},[],[2349],{"type":72,"value":1834},{"type":72,"value":2351}," — cycle-N+1 five-section leaderboard",{"type":66,"tag":216,"props":2353,"children":2354},{},[2355,2360,2362,2367],{"type":66,"tag":424,"props":2356,"children":2357},{},[2358],{"type":72,"value":2359},"(optional, after Step 4f)",{"type":72,"value":2361}," a deployed NIM endpoint (delegated to ",{"type":66,"tag":111,"props":2363,"children":2365},{"className":2364},[],[2366],{"type":72,"value":159},{"type":72,"value":2368},")",{"type":66,"tag":200,"props":2370,"children":2372},{"id":2371},"troubleshooting",[2373],{"type":72,"value":2374},"Troubleshooting",{"type":66,"tag":385,"props":2376,"children":2377},{},[2378,2411,2443,2467,2484,2506,2552],{"type":66,"tag":216,"props":2379,"children":2380},{},[2381,2386,2388,2394,2396,2401,2403,2409],{"type":66,"tag":83,"props":2382,"children":2383},{},[2384],{"type":72,"value":2385},"Stage 4 training collapses to all-UNK after first step",{"type":72,"value":2387}," → you're on the cache-aware streaming RNNT base (",{"type":66,"tag":111,"props":2389,"children":2391},{"className":2390},[],[2392],{"type":72,"value":2393},"nemotron-speech-streaming-en-0.6b",{"type":72,"value":2395},"). Route to ",{"type":66,"tag":111,"props":2397,"children":2399},{"className":2398},[],[2400],{"type":72,"value":282},{"type":72,"value":2402}," (the recommended default) or ",{"type":66,"tag":111,"props":2404,"children":2406},{"className":2405},[],[2407],{"type":72,"value":2408},"nvidia\u002Fstt_en_conformer_ctc_large",{"type":72,"value":2410}," (legacy fallback). The streaming RNNT SFT path is broken; do not retry with different hyperparameters.",{"type":66,"tag":216,"props":2412,"children":2413},{},[2414,2419,2421,2426,2428,2434,2436,2441],{"type":66,"tag":83,"props":2415,"children":2416},{},[2417],{"type":72,"value":2418},"Manifest paths don't resolve inside the NeMo container",{"type":72,"value":2420}," → host paths (e.g. ",{"type":66,"tag":111,"props":2422,"children":2424},{"className":2423},[],[2425],{"type":72,"value":1710},{"type":72,"value":2427},") need rewriting to ",{"type":66,"tag":111,"props":2429,"children":2431},{"className":2430},[],[2432],{"type":72,"value":2433},"\u002Fworkspace\u002F…",{"type":72,"value":2435},". See ",{"type":66,"tag":111,"props":2437,"children":2439},{"className":2438},[],[2440],{"type":72,"value":1223},{"type":72,"value":2442}," for the rewrite snippet.",{"type":66,"tag":216,"props":2444,"children":2445},{},[2446,2451,2453,2458,2460,2465],{"type":66,"tag":83,"props":2447,"children":2448},{},[2449],{"type":72,"value":2450},"Cycle N+1 KER unchanged from cycle N",{"type":72,"value":2452}," → on ",{"type":66,"tag":111,"props":2454,"children":2456},{"className":2455},[],[2457],{"type":72,"value":2194},{"type":72,"value":2459}," with the recipe above, this almost always means ",{"type":66,"tag":83,"props":2461,"children":2462},{},[2463],{"type":72,"value":2464},"manifest signal density is too low",{"type":72,"value":2466},". 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.)",{"type":66,"tag":216,"props":2468,"children":2469},{},[2470,2475,2477,2482],{"type":66,"tag":83,"props":2471,"children":2472},{},[2473],{"type":72,"value":2474},"Cycle N+1 KER got worse",{"type":72,"value":2476}," → overfit on a tiny manifest. Bail to ",{"type":66,"tag":111,"props":2478,"children":2480},{"className":2479},[],[2481],{"type":72,"value":325},{"type":72,"value":2483}," and grow.",{"type":66,"tag":216,"props":2485,"children":2486},{},[2487,2492,2494,2499,2500,2505],{"type":66,"tag":83,"props":2488,"children":2489},{},[2490],{"type":72,"value":2491},"Riva-served numbers diverge from offline numbers",{"type":72,"value":2493}," → the gap is in Riva preprocessing or ",{"type":66,"tag":111,"props":2495,"children":2497},{"className":2496},[],[2498],{"type":72,"value":2015},{"type":72,"value":2159},{"type":66,"tag":111,"props":2501,"children":2503},{"className":2502},[],[2504],{"type":72,"value":159},{"type":72,"value":266},{"type":66,"tag":216,"props":2507,"children":2508},{},[2509,2519,2521,2527,2529,2535,2537,2543,2545,2550],{"type":66,"tag":83,"props":2510,"children":2511},{},[2512,2517],{"type":66,"tag":111,"props":2513,"children":2515},{"className":2514},[],[2516],{"type":72,"value":563},{"type":72,"value":2518}," precision errors",{"type":72,"value":2520}," → some GPUs (older Turing, all Volta) don't support BF16. Drop to ",{"type":66,"tag":111,"props":2522,"children":2524},{"className":2523},[],[2525],{"type":72,"value":2526},"fp32",{"type":72,"value":2528}," and reduce ",{"type":66,"tag":111,"props":2530,"children":2532},{"className":2531},[],[2533],{"type":72,"value":2534},"batch_size",{"type":72,"value":2536},". Use ",{"type":66,"tag":111,"props":2538,"children":2540},{"className":2539},[],[2541],{"type":72,"value":2542},"fp16-mixed",{"type":72,"value":2544}," only if ",{"type":66,"tag":111,"props":2546,"children":2548},{"className":2547},[],[2549],{"type":72,"value":2526},{"type":72,"value":2551}," is too slow — fp16 with TDT decoders can produce NaN losses, so check loss curves early.",{"type":66,"tag":216,"props":2553,"children":2554},{},[2555,2560,2562,2567,2569,2575],{"type":66,"tag":83,"props":2556,"children":2557},{},[2558],{"type":72,"value":2559},"OOM during training on 24 GB GPU",{"type":72,"value":2561}," → drop ",{"type":66,"tag":111,"props":2563,"children":2565},{"className":2564},[],[2566],{"type":72,"value":2534},{"type":72,"value":2568}," to 2, raise ",{"type":66,"tag":111,"props":2570,"children":2572},{"className":2571},[],[2573],{"type":72,"value":2574},"accumulate_grad_batches",{"type":72,"value":2576}," to 2 to keep the effective batch size constant.",{"type":66,"tag":200,"props":2578,"children":2580},{"id":2579},"limitations",[2581],{"type":72,"value":2582},"Limitations",{"type":66,"tag":385,"props":2584,"children":2585},{},[2586,2610,2626,2636,2646],{"type":66,"tag":216,"props":2587,"children":2588},{},[2589,2594,2596,2601,2603,2608],{"type":66,"tag":83,"props":2590,"children":2591},{},[2592],{"type":72,"value":2593},"Adapter-style SFT on TDT\u002FRNNT decoders is broken.",{"type":72,"value":2595}," 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 ",{"type":66,"tag":83,"props":2597,"children":2598},{},[2599],{"type":72,"value":2600},"stock full-model SFT",{"type":72,"value":2602}," (",{"type":66,"tag":111,"props":2604,"children":2606},{"className":2605},[],[2607],{"type":72,"value":1480},{"type":72,"value":2609},") — which is what this skill recommends. Do not attempt adapter SFT on TDT\u002FRNNT bases.",{"type":66,"tag":216,"props":2611,"children":2612},{},[2613,2624],{"type":66,"tag":83,"props":2614,"children":2615},{},[2616,2618,2623],{"type":72,"value":2617},"Don't SFT ",{"type":66,"tag":111,"props":2619,"children":2621},{"className":2620},[],[2622],{"type":72,"value":2393},{"type":72,"value":266},{"type":72,"value":2625}," The streaming-only NVCF function's SFT path is unreliable (UNK collapse). For streaming serving at deploy time, Riva chunks a non-streaming base.",{"type":66,"tag":216,"props":2627,"children":2628},{},[2629,2634],{"type":66,"tag":83,"props":2630,"children":2631},{},[2632],{"type":72,"value":2633},"Tiny manifests overfit fast.",{"type":72,"value":2635}," Below ~100 rows total or ~5 rows per priority category, cycle-N+1 numbers are noisy. Grow before trusting a small KER drop.",{"type":66,"tag":216,"props":2637,"children":2638},{},[2639,2644],{"type":66,"tag":83,"props":2640,"children":2641},{},[2642],{"type":72,"value":2643},"English-only by default.",{"type":72,"value":2645}," The base-model table is en-US-specific. Other locales need a different base + a re-validated SFT recipe.",{"type":66,"tag":216,"props":2647,"children":2648},{},[2649,2654],{"type":66,"tag":83,"props":2650,"children":2651},{},[2652],{"type":72,"value":2653},"No turn-key driver.",{"type":72,"value":2655}," 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.",{"type":66,"tag":200,"props":2657,"children":2659},{"id":2658},"next-steps",[2660],{"type":72,"value":2661},"Next steps",{"type":66,"tag":385,"props":2663,"children":2664},{},[2665,2689,2704,2727],{"type":66,"tag":216,"props":2666,"children":2667},{},[2668,2680,2682,2687],{"type":66,"tag":83,"props":2669,"children":2670},{},[2671,2673,2678],{"type":72,"value":2672},"Deploy the ",{"type":66,"tag":111,"props":2674,"children":2676},{"className":2675},[],[2677],{"type":72,"value":151},{"type":72,"value":2679}," as a NIM:",{"type":72,"value":2681}," ",{"type":66,"tag":111,"props":2683,"children":2685},{"className":2684},[],[2686],{"type":72,"value":159},{"type":72,"value":2688}," (pass the source architecture explicitly).",{"type":66,"tag":216,"props":2690,"children":2691},{},[2692,2697,2698,2703],{"type":66,"tag":83,"props":2693,"children":2694},{},[2695],{"type":72,"value":2696},"Grow the manifest for cycle N+2:",{"type":72,"value":2681},{"type":66,"tag":111,"props":2699,"children":2701},{"className":2700},[],[2702],{"type":72,"value":325},{"type":72,"value":266},{"type":66,"tag":216,"props":2705,"children":2706},{},[2707,2712,2713,2718,2720,2725],{"type":66,"tag":83,"props":2708,"children":2709},{},[2710],{"type":72,"value":2711},"Re-score the cycle:",{"type":72,"value":2681},{"type":66,"tag":111,"props":2714,"children":2716},{"className":2715},[],[2717],{"type":72,"value":136},{"type":72,"value":2719}," (against the new endpoint or the new ",{"type":66,"tag":111,"props":2721,"children":2723},{"className":2722},[],[2724],{"type":72,"value":151},{"type":72,"value":2726}," directly).",{"type":66,"tag":216,"props":2728,"children":2729},{},[2730,2735,2737,2742],{"type":66,"tag":83,"props":2731,"children":2732},{},[2733],{"type":72,"value":2734},"Lateral",{"type":72,"value":2736}," for word boosting \u002F LM fusion \u002F non-clinical SFT recipes: ",{"type":66,"tag":111,"props":2738,"children":2740},{"className":2739},[],[2741],{"type":72,"value":488},{"type":72,"value":266},{"type":66,"tag":200,"props":2744,"children":2746},{"id":2745},"references",[2747],{"type":72,"value":2748},"References",{"type":66,"tag":385,"props":2750,"children":2751},{},[2752,2766],{"type":66,"tag":216,"props":2753,"children":2754},{},[2755,2764],{"type":66,"tag":2756,"props":2757,"children":2758},"a",{"href":116},[2759],{"type":66,"tag":111,"props":2760,"children":2762},{"className":2761},[],[2763],{"type":72,"value":116},{"type":72,"value":2765}," — base-model selection table, hyperparameter rationale, decoder → NIM container mapping, decision tree comparing cycle-N+1 to cycle-N",{"type":66,"tag":216,"props":2767,"children":2768},{},[2769,2777,2779,2785],{"type":66,"tag":2756,"props":2770,"children":2771},{"href":1223},[2772],{"type":66,"tag":111,"props":2773,"children":2775},{"className":2774},[],[2776],{"type":72,"value":1223},{"type":72,"value":2778}," — host → ",{"type":66,"tag":111,"props":2780,"children":2782},{"className":2781},[],[2783],{"type":72,"value":2784},"\u002Fworkspace\u002F",{"type":72,"value":2786}," path rewriting for cross-host manifest portability (laptop ↔ Brev ↔ NeMo container)",{"type":66,"tag":2788,"props":2789,"children":2790},"style",{},[2791],{"type":72,"value":2792},"html .light .shiki span {color: var(--shiki-light);background: var(--shiki-light-bg);font-style: var(--shiki-light-font-style);font-weight: var(--shiki-light-font-weight);text-decoration: var(--shiki-light-text-decoration);}html.light .shiki span {color: var(--shiki-light);background: var(--shiki-light-bg);font-style: var(--shiki-light-font-style);font-weight: var(--shiki-light-font-weight);text-decoration: var(--shiki-light-text-decoration);}html .default .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}html.dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}",{"items":2794,"total":2950},[2795,2813,2830,2841,2853,2867,2880,2892,2905,2916,2930,2939],{"slug":2796,"name":2796,"fn":2797,"description":2798,"org":2799,"tags":2800,"stars":2810,"repoUrl":2811,"updatedAt":2812},"nemoclaw-user-guide","retrieve NemoClaw documentation and configuration","Guides human users' AI agents to the NemoClaw docs MCP server and canonical Fern documentation in Markdown form. Use when users ask how to install, configure, operate, troubleshoot, secure, or learn NemoClaw with an AI coding assistant. Trigger keywords - nemoclaw docs, use nemoclaw with ai agent, nemoclaw mcp docs, nemoclaw install help, nemoclaw quickstart, nemoclaw markdown docs, llms.txt, agent skills.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[2801,2804,2807],{"name":2802,"slug":2803,"type":15},"Documentation","documentation",{"name":2805,"slug":2806,"type":15},"MCP","mcp",{"name":2808,"slug":2809,"type":15},"Search","search",21777,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FNemoClaw","2026-07-20T06:00:01.461044",{"slug":2814,"name":2814,"fn":2815,"description":2816,"org":2817,"tags":2818,"stars":2827,"repoUrl":2828,"updatedAt":2829},"mcore-build-and-dependency","manage Megatron-LM development environments","Container-based dev environment setup and dependency management for Megatron-LM. Covers acquiring and launching the CI container, uv package management, and updating uv.lock.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[2819,2822,2825],{"name":2820,"slug":2821,"type":15},"Containers","containers",{"name":2823,"slug":2824,"type":15},"Deployment","deployment",{"name":2826,"slug":1290,"type":15},"Python",17049,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FMegatron-LM","2026-07-27T06:06:11.249662",{"slug":2831,"name":2831,"fn":2832,"description":2833,"org":2834,"tags":2835,"stars":2827,"repoUrl":2828,"updatedAt":2840},"mcore-bump-base-image","update NVIDIA PyTorch base images","Bump the NVIDIA PyTorch base image (`nvcr.io\u002Fnvidia\u002Fpytorch:YY.MM-py3`) used by Megatron-LM CI. Covers the two pin sites (GitHub CI in `docker\u002F.ngc_version.dev` and GitLab CI in `.gitlab\u002Fstages\u002F01.build.yml`), the post-bump CI loop (re-run functional tests, refresh golden values, mark broken tests), and the gotchas that bit PRs",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[2836,2839],{"name":2837,"slug":2838,"type":15},"CI\u002FCD","ci-cd",{"name":2823,"slug":2824,"type":15},"2026-07-14T05:25:59.97109",{"slug":2842,"name":2842,"fn":2843,"description":2844,"org":2845,"tags":2846,"stars":2827,"repoUrl":2828,"updatedAt":2852},"mcore-cicd","manage CI\u002FCD pipelines for Megatron-LM","CI\u002FCD reference for Megatron-LM. Covers CI pipeline structure, PR scope labels, triggering internal GitLab CI (which force-pushes the current branch to a pull-request\u002FBRANCH ref — always dry-run and verify the destination first; never run against shared or protected branches), and CI failure investigation.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[2847,2848,2849],{"name":2837,"slug":2838,"type":15},{"name":2823,"slug":2824,"type":15},{"name":2850,"slug":2851,"type":15},"GitHub","github","2026-07-27T06:06:12.278222",{"slug":2854,"name":2854,"fn":2855,"description":2856,"org":2857,"tags":2858,"stars":2827,"repoUrl":2828,"updatedAt":2866},"mcore-create-issue","investigate CI failures and create issues","Investigate a failing GitHub Actions run or job and create a GitHub issue for the failure.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[2859,2862,2863],{"name":2860,"slug":2861,"type":15},"Debugging","debugging",{"name":2850,"slug":2851,"type":15},{"name":2864,"slug":2865,"type":15},"Triage","triage","2026-07-14T05:25:57.442089",{"slug":2868,"name":2868,"fn":2869,"description":2870,"org":2871,"tags":2872,"stars":2827,"repoUrl":2828,"updatedAt":2879},"mcore-linting-and-formatting","lint and format Megatron-LM code","Linting and formatting for Megatron-LM. Covers running autoformat.sh, tools (ruff, black, isort, pylint, mypy), and code style rules.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[2873,2876],{"name":2874,"slug":2875,"type":15},"Best Practices","best-practices",{"name":2877,"slug":2878,"type":15},"Code Analysis","code-analysis","2026-07-14T05:25:56.18433",{"slug":2881,"name":2881,"fn":2882,"description":2883,"org":2884,"tags":2885,"stars":2827,"repoUrl":2828,"updatedAt":2891},"mcore-migrate-gpt-to-hybrid","migrate Megatron-LM models to HybridModel","Migration guide for moving Megatron Core GPTModel checkpoints, model providers, training commands, and layer mappings to HybridModel.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[2886,2887,2890],{"name":17,"slug":18,"type":15},{"name":2888,"slug":2889,"type":15},"Migration","migration",{"name":9,"slug":8,"type":15},"2026-07-17T06:07:11.777011",{"slug":2893,"name":2893,"fn":2894,"description":2895,"org":2896,"tags":2897,"stars":2827,"repoUrl":2828,"updatedAt":2904},"mcore-onboard-gb200-1node-tests","onboard functional tests for GB200","Onboard 1-node GitHub MR functional tests for GB200 from existing mr-scoped 2-node tests.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[2898,2901],{"name":2899,"slug":2900,"type":15},"QA","qa",{"name":2902,"slug":2903,"type":15},"Testing","testing","2026-07-14T05:25:53.673039",{"slug":2906,"name":2906,"fn":2907,"description":2908,"org":2909,"tags":2910,"stars":2827,"repoUrl":2828,"updatedAt":2915},"mcore-run-on-slurm","launch distributed training jobs on SLURM","How to launch distributed Megatron-LM training jobs on a SLURM cluster. Covers a minimal sbatch skeleton, environment-variable setup for torch.distributed.run, CUDA_DEVICE_MAX_CONNECTIONS rules across hardware and parallelism modes, container conventions, monitoring, and per-rank failure diagnosis.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[2911,2912],{"name":2823,"slug":2824,"type":15},{"name":2913,"slug":2914,"type":15},"Infrastructure","infrastructure","2026-07-14T05:25:49.362534",{"slug":2917,"name":2917,"fn":2918,"description":2919,"org":2920,"tags":2921,"stars":2827,"repoUrl":2828,"updatedAt":2929},"mcore-split-pr","split pull requests to reduce review load","Split a PR into multiple PRs to reduce the number of required CODEOWNERS reviewer groups.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[2922,2925,2926],{"name":2923,"slug":2924,"type":15},"Code Review","code-review",{"name":2850,"slug":2851,"type":15},{"name":2927,"slug":2928,"type":15},"Pull Requests","pull-requests","2026-07-14T05:26:01.226578",{"slug":2931,"name":2931,"fn":2932,"description":2933,"org":2934,"tags":2935,"stars":2827,"repoUrl":2828,"updatedAt":2938},"mcore-testing","run and manage Megatron-LM tests","Test system for Megatron-LM. Covers test layout, recipe YAML structure, adding and running unit and functional tests, golden values, marker filters, and CI parity.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[2936,2937],{"name":2899,"slug":2900,"type":15},{"name":2902,"slug":2903,"type":15},"2026-07-14T05:25:54.928983",{"slug":2940,"name":2940,"fn":2941,"description":2942,"org":2943,"tags":2944,"stars":2827,"repoUrl":2828,"updatedAt":2949},"nightly-sync","manage nightly main-to-dev sync workflows","Domain knowledge for the nightly main-to-dev sync workflow. Covers merge strategy, CI architecture, failure investigation, and known issues.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[2945,2948],{"name":2946,"slug":2947,"type":15},"Automation","automation",{"name":2837,"slug":2838,"type":15},"2026-07-30T05:29:03.275638",496,{"items":2952,"total":3047},[2953,2970,2980,2994,3004,3019,3033],{"slug":2954,"name":2954,"fn":2955,"description":2956,"org":2957,"tags":2958,"stars":26,"repoUrl":27,"updatedAt":2969},"accelerated-computing-cudf","accelerate data processing with cuDF","Official NVIDIA-authored guidance for NVIDIA cuDF GPU DataFrames, pandas acceleration, dask-cuDF, ETL, joins, groupby, CSV\u002FParquet I\u002FO, nullable semantics, and multi-GPU DataFrame workloads.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[2959,2962,2965,2966],{"name":2960,"slug":2961,"type":15},"Data Analysis","data-analysis",{"name":2963,"slug":2964,"type":15},"Data Engineering","data-engineering",{"name":9,"slug":8,"type":15},{"name":2967,"slug":2968,"type":15},"Performance","performance","2026-07-14T05:28:43.176466",{"slug":2971,"name":2971,"fn":2972,"description":2973,"org":2974,"tags":2975,"stars":26,"repoUrl":27,"updatedAt":2979},"aiq-deploy","deploy and manage NVIDIA AI-Q infrastructure","Use when asked to install, deploy, run, validate, troubleshoot, or stop NVIDIA AI-Q Blueprint infrastructure.\n",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[2976,2977,2978],{"name":2823,"slug":2824,"type":15},{"name":2913,"slug":2914,"type":15},{"name":9,"slug":8,"type":15},"2026-07-14T05:29:06.667109",{"slug":2981,"name":2981,"fn":2982,"description":2983,"org":2984,"tags":2985,"stars":26,"repoUrl":27,"updatedAt":2993},"aiq-research","conduct deep research with AI-Q","Use when asked to run deep research or AI-Q research through a reachable NVIDIA AI-Q Blueprint backend.\n",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[2986,2989,2990],{"name":2987,"slug":2988,"type":15},"Agents","agents",{"name":9,"slug":8,"type":15},{"name":2991,"slug":2992,"type":15},"Research","research","2026-07-14T05:28:06.816956",{"slug":2995,"name":2995,"fn":2996,"description":2997,"org":2998,"tags":2999,"stars":26,"repoUrl":27,"updatedAt":3003},"amc-run-sample-calibration","run AMC sample dataset calibration","Run end-to-end calibration on the shipped sample dataset (sdg_08_2_sample_data_010926.zip) against a running AMC microservice. Use when user says 'test sample dataset', 'run sample calibration', 'verify AMC install', or 'launch and test'.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[3000,3001,3002],{"name":2960,"slug":2961,"type":15},{"name":9,"slug":8,"type":15},{"name":2902,"slug":2903,"type":15},"2026-07-17T05:29:03.913266",{"slug":3005,"name":3005,"fn":3006,"description":3007,"org":3008,"tags":3009,"stars":26,"repoUrl":27,"updatedAt":3018},"amc-run-video-calibration","calibrate video datasets with AutoMagicCalib","Calibrate a new dataset from pre-recorded video files via the AutoMagicCalib REST API. Use when user has local MP4s and says 'calibrate my videos', 'run AMC on these videos', or similar. For RTSP\u002Flive streams, use amc-run-rtsp-calibration instead.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[3010,3011,3014,3015],{"name":2946,"slug":2947,"type":15},{"name":3012,"slug":3013,"type":15},"Imaging","imaging",{"name":9,"slug":8,"type":15},{"name":3016,"slug":3017,"type":15},"Video","video","2026-07-17T05:28:53.905004",{"slug":3020,"name":3020,"fn":3021,"description":3022,"org":3023,"tags":3024,"stars":26,"repoUrl":27,"updatedAt":3032},"amc-setup-calibration-stack","deploy AutoMagicCalib microservice with Docker","Launch AutoMagicCalib microservice and web UI from NGC release images via Docker Compose. Use when user says 'deploy auto calibration', 'launch auto calibration', 'launch AMC', 'start MS+UI', or 'set up auto-magic-calib'. Requires NGC API key.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[3025,3026,3028,3029],{"name":2823,"slug":2824,"type":15},{"name":3027,"slug":1175,"type":15},"Docker",{"name":9,"slug":8,"type":15},{"name":3030,"slug":3031,"type":15},"Operations","operations","2026-07-17T05:28:56.913999",{"slug":3034,"name":3034,"fn":3035,"description":3036,"org":3037,"tags":3038,"stars":26,"repoUrl":27,"updatedAt":3046},"cudaq-guide","develop quantum applications with CUDA-Q","CUDA-Q onboarding guide for installation, test programs, GPU simulation, QPU hardware, and quantum applications.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[3039,3040,3043],{"name":9,"slug":8,"type":15},{"name":3041,"slug":3042,"type":15},"Quantum Computing","quantum-computing",{"name":3044,"slug":3045,"type":15},"Simulation","simulation","2026-07-14T05:26:58.898253",305]