[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill-nvidia-nemotron-asr-finetune":3,"mdc--kvdtsr-key":34,"related-org-nvidia-nemotron-asr-finetune":1134,"related-repo-nvidia-nemotron-asr-finetune":1292},{"slug":4,"name":4,"fn":5,"description":6,"org":7,"tags":11,"stars":23,"repoUrl":24,"updatedAt":25,"license":26,"forks":27,"topics":28,"repo":29,"sourceUrl":32,"mdContent":33},"nemotron-asr-finetune","fine-tune NVIDIA Nemotron ASR models","Orchestration skill for NVIDIA Nemotron Speech (Riva) \u002F NeMo ASR domain and language adaptation. Given a goal like \"improve\u002Ffine-tune ASR for my domain or language\", it scopes the task, picks the cheapest sufficient path (word boosting → n-gram LM → fine-tuning), delegates each stage to the right sub-skill (data generation, training, evaluation, deployment), and answers cost\u002Ftime\u002Fdata questions along the way.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},"nvidia","NVIDIA","https:\u002F\u002Fpexgzepcugksgbtrxkhf.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Forg-logos\u002Fnvidia.png",[12,16,19,20],{"name":13,"slug":14,"type":15},"Machine Learning","machine-learning","tag",{"name":17,"slug":18,"type":15},"Transcription","transcription",{"name":9,"slug":8,"type":15},{"name":21,"slug":22,"type":15},"Speech","speech",2473,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002Fskills","2026-07-20T06:24:01.073216","Apache-2.0",281,[],{"repoUrl":24,"stars":23,"forks":27,"topics":30,"description":31},[],"AI agent skills published by NVIDIA","https:\u002F\u002Fgithub.com\u002FNVIDIA\u002Fskills\u002Ftree\u002FHEAD\u002Fskills\u002Fnemotron-asr-finetune","---\nname: \"nemotron-asr-finetune\"\ndescription: Orchestration skill for NVIDIA Nemotron Speech (Riva) \u002F NeMo ASR domain and language adaptation. Given a goal like \"improve\u002Ffine-tune ASR for my domain or language\", it scopes the task, picks the cheapest sufficient path (word boosting → n-gram LM → fine-tuning), delegates each stage to the right sub-skill (data generation, training, evaluation, deployment), and answers cost\u002Ftime\u002Fdata questions along the way.\ntriggers:\n  - fine-tune ASR for my domain\n  - adapt ASR to my language\n  - improve ASR accuracy\n  - customize ASR\n  - ASR domain adaptation\n  - new language ASR\n  - reduce WER\n  - my ASR gets jargon wrong\n  - noisy call-center ASR\n  - how much data to fine-tune ASR\n  - cheapest way to improve ASR\n  - orchestrate ASR fine-tuning\n  - ASR customization pipeline\nversion: \"1.0.0\"\nlicense: Apache-2.0\nmetadata:\n  author: \"Nemotron Speech Team\"\n  team: riva\n  tags:\n    - nvidia\n    - nemotron-speech\n    - riva\n    - nemo\n    - asr\n    - speech-to-text\n    - orchestration\n    - customization\n    - domain-adaptation\n    - fine-tuning\n    - word-boosting\n    - language-model\n    - synthetic-data\n    - evaluation\n    - deployment\n  domain: ml\n---\n\n# Nemotron Speech ASR Customization — Orchestration Skill\n\n> **Note:** \"Nemotron Speech\" is the public-facing name for what NVIDIA documents today as **Riva** \u002F **Riva NIM**; the acoustic models are trained and fine-tuned with **NVIDIA NeMo**. Commands, config paths, imports, and doc URLs still use **\"Riva\"** \u002F **\"NeMo\"** — the rename is brand-only. Do not rename them.\n\n## What This Skill Is\n\nThis is a **high-level orchestration skill**, not a step-by-step training manual. Its job, given a goal such as *\"I want to fine-tune ASR for my domain\u002Flanguage\"*, is to:\n\n1. **Scope** the problem (how much real audio, target eval set, latency\u002Fhardware budget, language\u002Fdomain).\n2. **Choose the cheapest sufficient path** — word boosting, n-gram LM fusion, or fine-tuning — and escalate only when quality falls short.\n3. **Delegate each stage to the right sub-skill** (data generation, training, evaluation, deployment\u002Foptimization).\n4. **Answer cost\u002Ftime\u002Fdata questions along the way** (how many hours to hit X% WER, synthetic vs real, L40S vs H100, expected cost).\n\nIt owns the plan and the routing; the sub-skills own the execution. When a needed sub-skill does not exist yet, this skill names it as a **placeholder** and gives interim guidance.\n\n## When to Use\n\nUse for any request to make a Nemotron Speech \u002F Riva ASR model work better on a specific domain or language — improving accuracy, reducing WER, adding a language, or planning a fine-tune. Start here even when the user names a specific technique, so the cheapest sufficient path is chosen and the right sub-skills are sequenced.\n\n## Orchestration Workflow\n\nRun the loop below; each stage names the sub-skill it invokes. Full detail in [`references\u002Fworkflow.md`](references\u002Fworkflow.md).\n\n| # | Stage | What happens | Sub-skill |\n|---|---|---|---|\n| 1 | **State the goal** | Capture the target: domain\u002Flanguage, the errors, the metric. | Orchestration (this skill) |\n| 2 | **Clarify & scope** | Ask the discovery questions: how much real audio? target eval set? latency\u002FHW budget? deployment target? | Orchestration |\n| 3 | **Choose the path** | Pick the cheapest sufficient rung (boosting → n-gram LM → fine-tune). Escalate only if quality is short; experiment while proposing the full plan. | Orchestration → Research\u002FTraining |\n| 4 | **Get the data right** | If data is scarce\u002Fnoisy: synthetic (TTS), TTS-friendly formatting, noise profiling\u002Fharvest, blend, score vendor samples; align customer data to training format; flag missing real data. | SDG \u002F Data |\n| 5 | **Train** | Apply the recipe (configs, hyperparameters, replay\u002Fcurriculum, GPU\u002FOOM preflight) and run. | Research \u002F Training |\n| 6 | **Evaluate** | Normalized WER on the domain set + A\u002FB forgetting check on a general set; error-driven analysis to find the next lever. | Evaluation |\n| 7 | **Loop or ship** | If short of target, loop to 4\u002F5 with targeted data; else select\u002Faverage checkpoints. Consult the user before more cycles. | Orchestration |\n| 8 | **Deploy** | Export to NIM\u002FHF, hot-swap the checkpoint, serve. | Deployment \u002F Optimization |\n\n**Stages 4–8 are the fine-tune path** (`data → NeMo train → NeMo eval → Riva deploy`). Cheaper rungs (boosting, custom vocab, n-gram LM) take a **shorter branch owned by a single sub-skill** — don't force them through the full loop. See the branch-by-rung table in [`references\u002Fworkflow.md`](references\u002Fworkflow.md) (§3b).\n\nThroughout, answer the **\"along the way\"** questions (data volume, synthetic vs real, hours to reach a WER target, cost, GPU choice) — see [`references\u002Fplanning-answers.md`](references\u002Fplanning-answers.md).\n\n## Sub-Skills This Skill Calls\n\nDetailed registry, invocation, and handoff contracts in [`references\u002Fsub-skills.md`](references\u002Fsub-skills.md).\n\n| Role (per the architecture) | Purpose | Sub-skill to invoke |\n|---|---|---|\n| **Research \u002F Training** | NeMo configs, recipes, fine-tuning, checkpoint averaging | `nemo-speech-asr-finetune` |\n| **SDG \u002F Data Designer** | Synthetic transcripts\u002Ftext, noise profiling, vendor-data impact, blends | `data-designer` (synthetic **text**; audio via TTS in `nemotron-speech`); *placeholder:* `asr-data-profiling` |\n| **Evaluation** | Normalized WER, A\u002FB forgetting, error analysis | Offline file WER → `nemo-speech-asr-finetune`; **served-endpoint WER → `nemotron-speech`** |\n| **Deployment \u002F Optimization** | NIM\u002FRiva export, checkpoint swap, NIM-build optimization, serving | `nemotron-speech` |\n\nIf a sub-skill is unavailable, say so, give the interim guidance from the reference, and continue the plan.\n\n## Choosing The Path (cheapest first)\n\nThe scoping in Stage 3 selects the lowest-cost rung that can meet the target. Summary; full docs-grounded ladder in [`references\u002Fpath-selection.md`](references\u002Fpath-selection.md).\n\n- **Word boosting** — a bounded set of known words\u002Fnames\u002Fjargon. Runtime, no training. → Deployment sub-skill.\n- **Custom vocabulary \u002F pronunciation** — OOV or consistently mispronounced terms. Deploy-time. → Deployment sub-skill.\n- **N-gram (KenLM) LM** — domain phrasing\u002Fword-sequences when you have text but little audio. **Two realizations that are different artifacts:** *pilot (NeMo)* to prove lift offline (`nemo-speech-asr-finetune`), or *deploy (Riva)* to ship it (`nemotron-speech`). Don't ship the pilot LM — rebuild it in Riva word-level format. See [`references\u002Fpath-selection.md`](references\u002Fpath-selection.md).\n- **Fine-tune** — real acoustic gaps (accents, noise, channel) with enough transcribed audio (NIM guide: 100+ h; ~10 h floor only if mixed to avoid catastrophic forgetting). → Research\u002FTraining.\n- **Train from scratch \u002F cross-language transfer** — a new language with no suitable checkpoint (last resort). → Research\u002FTraining.\n\nOrdering and per-model support follow the NVIDIA Speech NIM ASR customization guide:\n\u003Chttps:\u002F\u002Fdocs.nvidia.com\u002Fnim\u002Fspeech\u002Flatest\u002Fasr\u002Fcustomization\u002Fcustomization.html>.\n\n## Key Principles\n\n- **Scope before you pick.** Don't recommend fine-tuning before the discovery questions and a measured baseline.\n- **Cheapest sufficient path.** Escalate rungs only when the current one provably can't hit the target; you may experiment on a cheap rung while presenting the full fine-tuning plan.\n- **Measure with a contract.** Report normalized WER on the domain set plus an A\u002FB forgetting check on a general set — never in-training logs alone.\n- **Delegate, don't reimplement.** Route execution to the sub-skills; keep this skill focused on the plan, sequencing, and cost\u002Ftime\u002Fdata answers.\n- **Real target-domain audio is the usual bottleneck.** Prefer real data; use synthetic to fill measured gaps, kept separately weighted so it can be ablated.\n- **Consult the user before extra tuning cycles**, and when a needed sub-skill is a placeholder.\n\n## Source of Truth\n\n| Topic | Location |\n|---|---|\n| NIM Speech docs home | https:\u002F\u002Fdocs.nvidia.com\u002Fnim\u002Fspeech\u002Flatest\u002Findex.html |\n| ASR customization guide (methods, per-model support) | https:\u002F\u002Fdocs.nvidia.com\u002Fnim\u002Fspeech\u002Flatest\u002Fasr\u002Fcustomization\u002Fcustomization.html |\n| ASR support matrix (models & features) | https:\u002F\u002Fdocs.nvidia.com\u002Fnim\u002Fspeech\u002Flatest\u002Freference\u002Fsupport-matrix\u002Fasr.html |\n| NeMo fine-tuning (flags\u002Fconfig) | `docs\u002Fsource\u002Fasr\u002Ffine_tuning.rst`, and the `nemo-speech-asr-finetune` sub-skill |\n| Riva ASR tutorials (boosting, LM, fine-tune) | https:\u002F\u002Fgithub.com\u002Fnvidia-riva\u002Ftutorials |\n| Tokenizer extension to new language + acoustic fine-tune | https:\u002F\u002Fgithub.com\u002Fnvidia-riva\u002Ftutorials\u002Fblob\u002Fmain\u002Fasr-extend-tokenizer-to-newlang-ft-acoustic-model.ipynb |\n\n## Limitations\n\n- Orchestration only — execution happens in the sub-skills. Where a sub-skill is a placeholder, guidance is interim until it exists.\n- GPU required for the training rungs; deployment\u002Fserving is owned by the `nemotron-speech` sub-skill.\n- Model names, config paths, flags, and per-model feature support drift across NeMo\u002FRiva releases — verify against the support matrix and the current checkout.\n- Public branding is **\"Nemotron Speech\"**; commands, imports, config paths, and doc URLs still use **\"Riva\"** \u002F **\"NeMo\"** — do not rename.\n",{"data":35,"body":69},{"name":4,"description":6,"triggers":36,"version":50,"license":26,"metadata":51},[37,38,39,40,41,42,43,44,45,46,47,48,49],"fine-tune ASR for my domain","adapt ASR to my language","improve ASR accuracy","customize ASR","ASR domain adaptation","new language ASR","reduce WER","my ASR gets jargon wrong","noisy call-center ASR","how much data to fine-tune ASR","cheapest way to improve ASR","orchestrate ASR fine-tuning","ASR customization pipeline","1.0.0",{"author":52,"team":53,"tags":54,"domain":68},"Nemotron Speech Team","riva",[8,55,53,56,57,58,59,60,61,62,63,64,65,66,67],"nemotron-speech","nemo","asr","speech-to-text","orchestration","customization","domain-adaptation","fine-tuning","word-boosting","language-model","synthetic-data","evaluation","deployment","ml",{"type":70,"children":71},"root",[72,81,131,138,158,203,215,221,226,232,250,495,530,552,558,573,737,742,748,763,860,873,879,942,948,1080,1086],{"type":73,"tag":74,"props":75,"children":77},"element","h1",{"id":76},"nemotron-speech-asr-customization-orchestration-skill",[78],{"type":79,"value":80},"text","Nemotron Speech ASR Customization — Orchestration Skill",{"type":73,"tag":82,"props":83,"children":84},"blockquote",{},[85],{"type":73,"tag":86,"props":87,"children":88},"p",{},[89,95,97,102,104,109,111,116,118,123,124,129],{"type":73,"tag":90,"props":91,"children":92},"strong",{},[93],{"type":79,"value":94},"Note:",{"type":79,"value":96}," \"Nemotron Speech\" is the public-facing name for what NVIDIA documents today as ",{"type":73,"tag":90,"props":98,"children":99},{},[100],{"type":79,"value":101},"Riva",{"type":79,"value":103}," \u002F ",{"type":73,"tag":90,"props":105,"children":106},{},[107],{"type":79,"value":108},"Riva NIM",{"type":79,"value":110},"; the acoustic models are trained and fine-tuned with ",{"type":73,"tag":90,"props":112,"children":113},{},[114],{"type":79,"value":115},"NVIDIA NeMo",{"type":79,"value":117},". Commands, config paths, imports, and doc URLs still use ",{"type":73,"tag":90,"props":119,"children":120},{},[121],{"type":79,"value":122},"\"Riva\"",{"type":79,"value":103},{"type":73,"tag":90,"props":125,"children":126},{},[127],{"type":79,"value":128},"\"NeMo\"",{"type":79,"value":130}," — the rename is brand-only. Do not rename them.",{"type":73,"tag":132,"props":133,"children":135},"h2",{"id":134},"what-this-skill-is",[136],{"type":79,"value":137},"What This Skill Is",{"type":73,"tag":86,"props":139,"children":140},{},[141,143,148,150,156],{"type":79,"value":142},"This is a ",{"type":73,"tag":90,"props":144,"children":145},{},[146],{"type":79,"value":147},"high-level orchestration skill",{"type":79,"value":149},", not a step-by-step training manual. Its job, given a goal such as ",{"type":73,"tag":151,"props":152,"children":153},"em",{},[154],{"type":79,"value":155},"\"I want to fine-tune ASR for my domain\u002Flanguage\"",{"type":79,"value":157},", is to:",{"type":73,"tag":159,"props":160,"children":161},"ol",{},[162,173,183,193],{"type":73,"tag":163,"props":164,"children":165},"li",{},[166,171],{"type":73,"tag":90,"props":167,"children":168},{},[169],{"type":79,"value":170},"Scope",{"type":79,"value":172}," the problem (how much real audio, target eval set, latency\u002Fhardware budget, language\u002Fdomain).",{"type":73,"tag":163,"props":174,"children":175},{},[176,181],{"type":73,"tag":90,"props":177,"children":178},{},[179],{"type":79,"value":180},"Choose the cheapest sufficient path",{"type":79,"value":182}," — word boosting, n-gram LM fusion, or fine-tuning — and escalate only when quality falls short.",{"type":73,"tag":163,"props":184,"children":185},{},[186,191],{"type":73,"tag":90,"props":187,"children":188},{},[189],{"type":79,"value":190},"Delegate each stage to the right sub-skill",{"type":79,"value":192}," (data generation, training, evaluation, deployment\u002Foptimization).",{"type":73,"tag":163,"props":194,"children":195},{},[196,201],{"type":73,"tag":90,"props":197,"children":198},{},[199],{"type":79,"value":200},"Answer cost\u002Ftime\u002Fdata questions along the way",{"type":79,"value":202}," (how many hours to hit X% WER, synthetic vs real, L40S vs H100, expected cost).",{"type":73,"tag":86,"props":204,"children":205},{},[206,208,213],{"type":79,"value":207},"It owns the plan and the routing; the sub-skills own the execution. When a needed sub-skill does not exist yet, this skill names it as a ",{"type":73,"tag":90,"props":209,"children":210},{},[211],{"type":79,"value":212},"placeholder",{"type":79,"value":214}," and gives interim guidance.",{"type":73,"tag":132,"props":216,"children":218},{"id":217},"when-to-use",[219],{"type":79,"value":220},"When to Use",{"type":73,"tag":86,"props":222,"children":223},{},[224],{"type":79,"value":225},"Use for any request to make a Nemotron Speech \u002F Riva ASR model work better on a specific domain or language — improving accuracy, reducing WER, adding a language, or planning a fine-tune. Start here even when the user names a specific technique, so the cheapest sufficient path is chosen and the right sub-skills are sequenced.",{"type":73,"tag":132,"props":227,"children":229},{"id":228},"orchestration-workflow",[230],{"type":79,"value":231},"Orchestration Workflow",{"type":73,"tag":86,"props":233,"children":234},{},[235,237,248],{"type":79,"value":236},"Run the loop below; each stage names the sub-skill it invokes. Full detail in ",{"type":73,"tag":238,"props":239,"children":241},"a",{"href":240},"references\u002Fworkflow.md",[242],{"type":73,"tag":243,"props":244,"children":246},"code",{"className":245},[],[247],{"type":79,"value":240},{"type":79,"value":249},".",{"type":73,"tag":251,"props":252,"children":253},"table",{},[254,283],{"type":73,"tag":255,"props":256,"children":257},"thead",{},[258],{"type":73,"tag":259,"props":260,"children":261},"tr",{},[262,268,273,278],{"type":73,"tag":263,"props":264,"children":265},"th",{},[266],{"type":79,"value":267},"#",{"type":73,"tag":263,"props":269,"children":270},{},[271],{"type":79,"value":272},"Stage",{"type":73,"tag":263,"props":274,"children":275},{},[276],{"type":79,"value":277},"What happens",{"type":73,"tag":263,"props":279,"children":280},{},[281],{"type":79,"value":282},"Sub-skill",{"type":73,"tag":284,"props":285,"children":286},"tbody",{},[287,314,340,366,392,418,444,469],{"type":73,"tag":259,"props":288,"children":289},{},[290,296,304,309],{"type":73,"tag":291,"props":292,"children":293},"td",{},[294],{"type":79,"value":295},"1",{"type":73,"tag":291,"props":297,"children":298},{},[299],{"type":73,"tag":90,"props":300,"children":301},{},[302],{"type":79,"value":303},"State the goal",{"type":73,"tag":291,"props":305,"children":306},{},[307],{"type":79,"value":308},"Capture the target: domain\u002Flanguage, the errors, the metric.",{"type":73,"tag":291,"props":310,"children":311},{},[312],{"type":79,"value":313},"Orchestration (this skill)",{"type":73,"tag":259,"props":315,"children":316},{},[317,322,330,335],{"type":73,"tag":291,"props":318,"children":319},{},[320],{"type":79,"value":321},"2",{"type":73,"tag":291,"props":323,"children":324},{},[325],{"type":73,"tag":90,"props":326,"children":327},{},[328],{"type":79,"value":329},"Clarify & scope",{"type":73,"tag":291,"props":331,"children":332},{},[333],{"type":79,"value":334},"Ask the discovery questions: how much real audio? target eval set? latency\u002FHW budget? deployment target?",{"type":73,"tag":291,"props":336,"children":337},{},[338],{"type":79,"value":339},"Orchestration",{"type":73,"tag":259,"props":341,"children":342},{},[343,348,356,361],{"type":73,"tag":291,"props":344,"children":345},{},[346],{"type":79,"value":347},"3",{"type":73,"tag":291,"props":349,"children":350},{},[351],{"type":73,"tag":90,"props":352,"children":353},{},[354],{"type":79,"value":355},"Choose the path",{"type":73,"tag":291,"props":357,"children":358},{},[359],{"type":79,"value":360},"Pick the cheapest sufficient rung (boosting → n-gram LM → fine-tune). Escalate only if quality is short; experiment while proposing the full plan.",{"type":73,"tag":291,"props":362,"children":363},{},[364],{"type":79,"value":365},"Orchestration → Research\u002FTraining",{"type":73,"tag":259,"props":367,"children":368},{},[369,374,382,387],{"type":73,"tag":291,"props":370,"children":371},{},[372],{"type":79,"value":373},"4",{"type":73,"tag":291,"props":375,"children":376},{},[377],{"type":73,"tag":90,"props":378,"children":379},{},[380],{"type":79,"value":381},"Get the data right",{"type":73,"tag":291,"props":383,"children":384},{},[385],{"type":79,"value":386},"If data is scarce\u002Fnoisy: synthetic (TTS), TTS-friendly formatting, noise profiling\u002Fharvest, blend, score vendor samples; align customer data to training format; flag missing real data.",{"type":73,"tag":291,"props":388,"children":389},{},[390],{"type":79,"value":391},"SDG \u002F Data",{"type":73,"tag":259,"props":393,"children":394},{},[395,400,408,413],{"type":73,"tag":291,"props":396,"children":397},{},[398],{"type":79,"value":399},"5",{"type":73,"tag":291,"props":401,"children":402},{},[403],{"type":73,"tag":90,"props":404,"children":405},{},[406],{"type":79,"value":407},"Train",{"type":73,"tag":291,"props":409,"children":410},{},[411],{"type":79,"value":412},"Apply the recipe (configs, hyperparameters, replay\u002Fcurriculum, GPU\u002FOOM preflight) and run.",{"type":73,"tag":291,"props":414,"children":415},{},[416],{"type":79,"value":417},"Research \u002F Training",{"type":73,"tag":259,"props":419,"children":420},{},[421,426,434,439],{"type":73,"tag":291,"props":422,"children":423},{},[424],{"type":79,"value":425},"6",{"type":73,"tag":291,"props":427,"children":428},{},[429],{"type":73,"tag":90,"props":430,"children":431},{},[432],{"type":79,"value":433},"Evaluate",{"type":73,"tag":291,"props":435,"children":436},{},[437],{"type":79,"value":438},"Normalized WER on the domain set + A\u002FB forgetting check on a general set; error-driven analysis to find the next lever.",{"type":73,"tag":291,"props":440,"children":441},{},[442],{"type":79,"value":443},"Evaluation",{"type":73,"tag":259,"props":445,"children":446},{},[447,452,460,465],{"type":73,"tag":291,"props":448,"children":449},{},[450],{"type":79,"value":451},"7",{"type":73,"tag":291,"props":453,"children":454},{},[455],{"type":73,"tag":90,"props":456,"children":457},{},[458],{"type":79,"value":459},"Loop or ship",{"type":73,"tag":291,"props":461,"children":462},{},[463],{"type":79,"value":464},"If short of target, loop to 4\u002F5 with targeted data; else select\u002Faverage checkpoints. Consult the user before more cycles.",{"type":73,"tag":291,"props":466,"children":467},{},[468],{"type":79,"value":339},{"type":73,"tag":259,"props":470,"children":471},{},[472,477,485,490],{"type":73,"tag":291,"props":473,"children":474},{},[475],{"type":79,"value":476},"8",{"type":73,"tag":291,"props":478,"children":479},{},[480],{"type":73,"tag":90,"props":481,"children":482},{},[483],{"type":79,"value":484},"Deploy",{"type":73,"tag":291,"props":486,"children":487},{},[488],{"type":79,"value":489},"Export to NIM\u002FHF, hot-swap the checkpoint, serve.",{"type":73,"tag":291,"props":491,"children":492},{},[493],{"type":79,"value":494},"Deployment \u002F Optimization",{"type":73,"tag":86,"props":496,"children":497},{},[498,503,505,511,513,518,520,528],{"type":73,"tag":90,"props":499,"children":500},{},[501],{"type":79,"value":502},"Stages 4–8 are the fine-tune path",{"type":79,"value":504}," (",{"type":73,"tag":243,"props":506,"children":508},{"className":507},[],[509],{"type":79,"value":510},"data → NeMo train → NeMo eval → Riva deploy",{"type":79,"value":512},"). Cheaper rungs (boosting, custom vocab, n-gram LM) take a ",{"type":73,"tag":90,"props":514,"children":515},{},[516],{"type":79,"value":517},"shorter branch owned by a single sub-skill",{"type":79,"value":519}," — don't force them through the full loop. See the branch-by-rung table in ",{"type":73,"tag":238,"props":521,"children":522},{"href":240},[523],{"type":73,"tag":243,"props":524,"children":526},{"className":525},[],[527],{"type":79,"value":240},{"type":79,"value":529}," (§3b).",{"type":73,"tag":86,"props":531,"children":532},{},[533,535,540,542,551],{"type":79,"value":534},"Throughout, answer the ",{"type":73,"tag":90,"props":536,"children":537},{},[538],{"type":79,"value":539},"\"along the way\"",{"type":79,"value":541}," questions (data volume, synthetic vs real, hours to reach a WER target, cost, GPU choice) — see ",{"type":73,"tag":238,"props":543,"children":545},{"href":544},"references\u002Fplanning-answers.md",[546],{"type":73,"tag":243,"props":547,"children":549},{"className":548},[],[550],{"type":79,"value":544},{"type":79,"value":249},{"type":73,"tag":132,"props":553,"children":555},{"id":554},"sub-skills-this-skill-calls",[556],{"type":79,"value":557},"Sub-Skills This Skill Calls",{"type":73,"tag":86,"props":559,"children":560},{},[561,563,572],{"type":79,"value":562},"Detailed registry, invocation, and handoff contracts in ",{"type":73,"tag":238,"props":564,"children":566},{"href":565},"references\u002Fsub-skills.md",[567],{"type":73,"tag":243,"props":568,"children":570},{"className":569},[],[571],{"type":79,"value":565},{"type":79,"value":249},{"type":73,"tag":251,"props":574,"children":575},{},[576,597],{"type":73,"tag":255,"props":577,"children":578},{},[579],{"type":73,"tag":259,"props":580,"children":581},{},[582,587,592],{"type":73,"tag":263,"props":583,"children":584},{},[585],{"type":79,"value":586},"Role (per the architecture)",{"type":73,"tag":263,"props":588,"children":589},{},[590],{"type":79,"value":591},"Purpose",{"type":73,"tag":263,"props":593,"children":594},{},[595],{"type":79,"value":596},"Sub-skill to invoke",{"type":73,"tag":284,"props":598,"children":599},{},[600,624,677,714],{"type":73,"tag":259,"props":601,"children":602},{},[603,610,615],{"type":73,"tag":291,"props":604,"children":605},{},[606],{"type":73,"tag":90,"props":607,"children":608},{},[609],{"type":79,"value":417},{"type":73,"tag":291,"props":611,"children":612},{},[613],{"type":79,"value":614},"NeMo configs, recipes, fine-tuning, checkpoint averaging",{"type":73,"tag":291,"props":616,"children":617},{},[618],{"type":73,"tag":243,"props":619,"children":621},{"className":620},[],[622],{"type":79,"value":623},"nemo-speech-asr-finetune",{"type":73,"tag":259,"props":625,"children":626},{},[627,635,640],{"type":73,"tag":291,"props":628,"children":629},{},[630],{"type":73,"tag":90,"props":631,"children":632},{},[633],{"type":79,"value":634},"SDG \u002F Data Designer",{"type":73,"tag":291,"props":636,"children":637},{},[638],{"type":79,"value":639},"Synthetic transcripts\u002Ftext, noise profiling, vendor-data impact, blends",{"type":73,"tag":291,"props":641,"children":642},{},[643,649,651,655,657,662,664,669,671],{"type":73,"tag":243,"props":644,"children":646},{"className":645},[],[647],{"type":79,"value":648},"data-designer",{"type":79,"value":650}," (synthetic ",{"type":73,"tag":90,"props":652,"children":653},{},[654],{"type":79,"value":79},{"type":79,"value":656},"; audio via TTS in ",{"type":73,"tag":243,"props":658,"children":660},{"className":659},[],[661],{"type":79,"value":55},{"type":79,"value":663},"); ",{"type":73,"tag":151,"props":665,"children":666},{},[667],{"type":79,"value":668},"placeholder:",{"type":79,"value":670}," ",{"type":73,"tag":243,"props":672,"children":674},{"className":673},[],[675],{"type":79,"value":676},"asr-data-profiling",{"type":73,"tag":259,"props":678,"children":679},{},[680,687,692],{"type":73,"tag":291,"props":681,"children":682},{},[683],{"type":73,"tag":90,"props":684,"children":685},{},[686],{"type":79,"value":443},{"type":73,"tag":291,"props":688,"children":689},{},[690],{"type":79,"value":691},"Normalized WER, A\u002FB forgetting, error analysis",{"type":73,"tag":291,"props":693,"children":694},{},[695,697,702,704],{"type":79,"value":696},"Offline file WER → ",{"type":73,"tag":243,"props":698,"children":700},{"className":699},[],[701],{"type":79,"value":623},{"type":79,"value":703},"; ",{"type":73,"tag":90,"props":705,"children":706},{},[707,709],{"type":79,"value":708},"served-endpoint WER → ",{"type":73,"tag":243,"props":710,"children":712},{"className":711},[],[713],{"type":79,"value":55},{"type":73,"tag":259,"props":715,"children":716},{},[717,724,729],{"type":73,"tag":291,"props":718,"children":719},{},[720],{"type":73,"tag":90,"props":721,"children":722},{},[723],{"type":79,"value":494},{"type":73,"tag":291,"props":725,"children":726},{},[727],{"type":79,"value":728},"NIM\u002FRiva export, checkpoint swap, NIM-build optimization, serving",{"type":73,"tag":291,"props":730,"children":731},{},[732],{"type":73,"tag":243,"props":733,"children":735},{"className":734},[],[736],{"type":79,"value":55},{"type":73,"tag":86,"props":738,"children":739},{},[740],{"type":79,"value":741},"If a sub-skill is unavailable, say so, give the interim guidance from the reference, and continue the plan.",{"type":73,"tag":132,"props":743,"children":745},{"id":744},"choosing-the-path-cheapest-first",[746],{"type":79,"value":747},"Choosing The Path (cheapest first)",{"type":73,"tag":86,"props":749,"children":750},{},[751,753,762],{"type":79,"value":752},"The scoping in Stage 3 selects the lowest-cost rung that can meet the target. Summary; full docs-grounded ladder in ",{"type":73,"tag":238,"props":754,"children":756},{"href":755},"references\u002Fpath-selection.md",[757],{"type":73,"tag":243,"props":758,"children":760},{"className":759},[],[761],{"type":79,"value":755},{"type":79,"value":249},{"type":73,"tag":764,"props":765,"children":766},"ul",{},[767,777,787,840,850],{"type":73,"tag":163,"props":768,"children":769},{},[770,775],{"type":73,"tag":90,"props":771,"children":772},{},[773],{"type":79,"value":774},"Word boosting",{"type":79,"value":776}," — a bounded set of known words\u002Fnames\u002Fjargon. Runtime, no training. → Deployment sub-skill.",{"type":73,"tag":163,"props":778,"children":779},{},[780,785],{"type":73,"tag":90,"props":781,"children":782},{},[783],{"type":79,"value":784},"Custom vocabulary \u002F pronunciation",{"type":79,"value":786}," — OOV or consistently mispronounced terms. Deploy-time. → Deployment sub-skill.",{"type":73,"tag":163,"props":788,"children":789},{},[790,795,797,802,803,808,810,815,817,822,824,829,831,839],{"type":73,"tag":90,"props":791,"children":792},{},[793],{"type":79,"value":794},"N-gram (KenLM) LM",{"type":79,"value":796}," — domain phrasing\u002Fword-sequences when you have text but little audio. ",{"type":73,"tag":90,"props":798,"children":799},{},[800],{"type":79,"value":801},"Two realizations that are different artifacts:",{"type":79,"value":670},{"type":73,"tag":151,"props":804,"children":805},{},[806],{"type":79,"value":807},"pilot (NeMo)",{"type":79,"value":809}," to prove lift offline (",{"type":73,"tag":243,"props":811,"children":813},{"className":812},[],[814],{"type":79,"value":623},{"type":79,"value":816},"), or ",{"type":73,"tag":151,"props":818,"children":819},{},[820],{"type":79,"value":821},"deploy (Riva)",{"type":79,"value":823}," to ship it (",{"type":73,"tag":243,"props":825,"children":827},{"className":826},[],[828],{"type":79,"value":55},{"type":79,"value":830},"). Don't ship the pilot LM — rebuild it in Riva word-level format. See ",{"type":73,"tag":238,"props":832,"children":833},{"href":755},[834],{"type":73,"tag":243,"props":835,"children":837},{"className":836},[],[838],{"type":79,"value":755},{"type":79,"value":249},{"type":73,"tag":163,"props":841,"children":842},{},[843,848],{"type":73,"tag":90,"props":844,"children":845},{},[846],{"type":79,"value":847},"Fine-tune",{"type":79,"value":849}," — real acoustic gaps (accents, noise, channel) with enough transcribed audio (NIM guide: 100+ h; ~10 h floor only if mixed to avoid catastrophic forgetting). → Research\u002FTraining.",{"type":73,"tag":163,"props":851,"children":852},{},[853,858],{"type":73,"tag":90,"props":854,"children":855},{},[856],{"type":79,"value":857},"Train from scratch \u002F cross-language transfer",{"type":79,"value":859}," — a new language with no suitable checkpoint (last resort). → Research\u002FTraining.",{"type":73,"tag":86,"props":861,"children":862},{},[863,865,872],{"type":79,"value":864},"Ordering and per-model support follow the NVIDIA Speech NIM ASR customization guide:\n",{"type":73,"tag":238,"props":866,"children":870},{"href":867,"rel":868},"https:\u002F\u002Fdocs.nvidia.com\u002Fnim\u002Fspeech\u002Flatest\u002Fasr\u002Fcustomization\u002Fcustomization.html",[869],"nofollow",[871],{"type":79,"value":867},{"type":79,"value":249},{"type":73,"tag":132,"props":874,"children":876},{"id":875},"key-principles",[877],{"type":79,"value":878},"Key Principles",{"type":73,"tag":764,"props":880,"children":881},{},[882,892,902,912,922,932],{"type":73,"tag":163,"props":883,"children":884},{},[885,890],{"type":73,"tag":90,"props":886,"children":887},{},[888],{"type":79,"value":889},"Scope before you pick.",{"type":79,"value":891}," Don't recommend fine-tuning before the discovery questions and a measured baseline.",{"type":73,"tag":163,"props":893,"children":894},{},[895,900],{"type":73,"tag":90,"props":896,"children":897},{},[898],{"type":79,"value":899},"Cheapest sufficient path.",{"type":79,"value":901}," Escalate rungs only when the current one provably can't hit the target; you may experiment on a cheap rung while presenting the full fine-tuning plan.",{"type":73,"tag":163,"props":903,"children":904},{},[905,910],{"type":73,"tag":90,"props":906,"children":907},{},[908],{"type":79,"value":909},"Measure with a contract.",{"type":79,"value":911}," Report normalized WER on the domain set plus an A\u002FB forgetting check on a general set — never in-training logs alone.",{"type":73,"tag":163,"props":913,"children":914},{},[915,920],{"type":73,"tag":90,"props":916,"children":917},{},[918],{"type":79,"value":919},"Delegate, don't reimplement.",{"type":79,"value":921}," Route execution to the sub-skills; keep this skill focused on the plan, sequencing, and cost\u002Ftime\u002Fdata answers.",{"type":73,"tag":163,"props":923,"children":924},{},[925,930],{"type":73,"tag":90,"props":926,"children":927},{},[928],{"type":79,"value":929},"Real target-domain audio is the usual bottleneck.",{"type":79,"value":931}," Prefer real data; use synthetic to fill measured gaps, kept separately weighted so it can be ablated.",{"type":73,"tag":163,"props":933,"children":934},{},[935,940],{"type":73,"tag":90,"props":936,"children":937},{},[938],{"type":79,"value":939},"Consult the user before extra tuning cycles",{"type":79,"value":941},", and when a needed sub-skill is a placeholder.",{"type":73,"tag":132,"props":943,"children":945},{"id":944},"source-of-truth",[946],{"type":79,"value":947},"Source of Truth",{"type":73,"tag":251,"props":949,"children":950},{},[951,967],{"type":73,"tag":255,"props":952,"children":953},{},[954],{"type":73,"tag":259,"props":955,"children":956},{},[957,962],{"type":73,"tag":263,"props":958,"children":959},{},[960],{"type":79,"value":961},"Topic",{"type":73,"tag":263,"props":963,"children":964},{},[965],{"type":79,"value":966},"Location",{"type":73,"tag":284,"props":968,"children":969},{},[970,987,1003,1020,1046,1063],{"type":73,"tag":259,"props":971,"children":972},{},[973,978],{"type":73,"tag":291,"props":974,"children":975},{},[976],{"type":79,"value":977},"NIM Speech docs home",{"type":73,"tag":291,"props":979,"children":980},{},[981],{"type":73,"tag":238,"props":982,"children":985},{"href":983,"rel":984},"https:\u002F\u002Fdocs.nvidia.com\u002Fnim\u002Fspeech\u002Flatest\u002Findex.html",[869],[986],{"type":79,"value":983},{"type":73,"tag":259,"props":988,"children":989},{},[990,995],{"type":73,"tag":291,"props":991,"children":992},{},[993],{"type":79,"value":994},"ASR customization guide (methods, per-model support)",{"type":73,"tag":291,"props":996,"children":997},{},[998],{"type":73,"tag":238,"props":999,"children":1001},{"href":867,"rel":1000},[869],[1002],{"type":79,"value":867},{"type":73,"tag":259,"props":1004,"children":1005},{},[1006,1011],{"type":73,"tag":291,"props":1007,"children":1008},{},[1009],{"type":79,"value":1010},"ASR support matrix (models & features)",{"type":73,"tag":291,"props":1012,"children":1013},{},[1014],{"type":73,"tag":238,"props":1015,"children":1018},{"href":1016,"rel":1017},"https:\u002F\u002Fdocs.nvidia.com\u002Fnim\u002Fspeech\u002Flatest\u002Freference\u002Fsupport-matrix\u002Fasr.html",[869],[1019],{"type":79,"value":1016},{"type":73,"tag":259,"props":1021,"children":1022},{},[1023,1028],{"type":73,"tag":291,"props":1024,"children":1025},{},[1026],{"type":79,"value":1027},"NeMo fine-tuning (flags\u002Fconfig)",{"type":73,"tag":291,"props":1029,"children":1030},{},[1031,1037,1039,1044],{"type":73,"tag":243,"props":1032,"children":1034},{"className":1033},[],[1035],{"type":79,"value":1036},"docs\u002Fsource\u002Fasr\u002Ffine_tuning.rst",{"type":79,"value":1038},", and the ",{"type":73,"tag":243,"props":1040,"children":1042},{"className":1041},[],[1043],{"type":79,"value":623},{"type":79,"value":1045}," sub-skill",{"type":73,"tag":259,"props":1047,"children":1048},{},[1049,1054],{"type":73,"tag":291,"props":1050,"children":1051},{},[1052],{"type":79,"value":1053},"Riva ASR tutorials (boosting, LM, fine-tune)",{"type":73,"tag":291,"props":1055,"children":1056},{},[1057],{"type":73,"tag":238,"props":1058,"children":1061},{"href":1059,"rel":1060},"https:\u002F\u002Fgithub.com\u002Fnvidia-riva\u002Ftutorials",[869],[1062],{"type":79,"value":1059},{"type":73,"tag":259,"props":1064,"children":1065},{},[1066,1071],{"type":73,"tag":291,"props":1067,"children":1068},{},[1069],{"type":79,"value":1070},"Tokenizer extension to new language + acoustic fine-tune",{"type":73,"tag":291,"props":1072,"children":1073},{},[1074],{"type":73,"tag":238,"props":1075,"children":1078},{"href":1076,"rel":1077},"https:\u002F\u002Fgithub.com\u002Fnvidia-riva\u002Ftutorials\u002Fblob\u002Fmain\u002Fasr-extend-tokenizer-to-newlang-ft-acoustic-model.ipynb",[869],[1079],{"type":79,"value":1076},{"type":73,"tag":132,"props":1081,"children":1083},{"id":1082},"limitations",[1084],{"type":79,"value":1085},"Limitations",{"type":73,"tag":764,"props":1087,"children":1088},{},[1089,1094,1106,1111],{"type":73,"tag":163,"props":1090,"children":1091},{},[1092],{"type":79,"value":1093},"Orchestration only — execution happens in the sub-skills. Where a sub-skill is a placeholder, guidance is interim until it exists.",{"type":73,"tag":163,"props":1095,"children":1096},{},[1097,1099,1104],{"type":79,"value":1098},"GPU required for the training rungs; deployment\u002Fserving is owned by the ",{"type":73,"tag":243,"props":1100,"children":1102},{"className":1101},[],[1103],{"type":79,"value":55},{"type":79,"value":1105}," sub-skill.",{"type":73,"tag":163,"props":1107,"children":1108},{},[1109],{"type":79,"value":1110},"Model names, config paths, flags, and per-model feature support drift across NeMo\u002FRiva releases — verify against the support matrix and the current checkout.",{"type":73,"tag":163,"props":1112,"children":1113},{},[1114,1116,1121,1123,1127,1128,1132],{"type":79,"value":1115},"Public branding is ",{"type":73,"tag":90,"props":1117,"children":1118},{},[1119],{"type":79,"value":1120},"\"Nemotron Speech\"",{"type":79,"value":1122},"; commands, imports, config paths, and doc URLs still use ",{"type":73,"tag":90,"props":1124,"children":1125},{},[1126],{"type":79,"value":122},{"type":79,"value":103},{"type":73,"tag":90,"props":1129,"children":1130},{},[1131],{"type":79,"value":128},{"type":79,"value":1133}," — do not rename.",{"items":1135,"total":1291},[1136,1154,1171,1182,1194,1208,1221,1233,1246,1257,1271,1280],{"slug":1137,"name":1137,"fn":1138,"description":1139,"org":1140,"tags":1141,"stars":1151,"repoUrl":1152,"updatedAt":1153},"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},[1142,1145,1148],{"name":1143,"slug":1144,"type":15},"Documentation","documentation",{"name":1146,"slug":1147,"type":15},"MCP","mcp",{"name":1149,"slug":1150,"type":15},"Search","search",21777,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FNemoClaw","2026-07-20T06:00:01.461044",{"slug":1155,"name":1155,"fn":1156,"description":1157,"org":1158,"tags":1159,"stars":1168,"repoUrl":1169,"updatedAt":1170},"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},[1160,1163,1165],{"name":1161,"slug":1162,"type":15},"Containers","containers",{"name":1164,"slug":67,"type":15},"Deployment",{"name":1166,"slug":1167,"type":15},"Python","python",17049,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FMegatron-LM","2026-07-27T06:06:11.249662",{"slug":1172,"name":1172,"fn":1173,"description":1174,"org":1175,"tags":1176,"stars":1168,"repoUrl":1169,"updatedAt":1181},"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},[1177,1180],{"name":1178,"slug":1179,"type":15},"CI\u002FCD","ci-cd",{"name":1164,"slug":67,"type":15},"2026-07-14T05:25:59.97109",{"slug":1183,"name":1183,"fn":1184,"description":1185,"org":1186,"tags":1187,"stars":1168,"repoUrl":1169,"updatedAt":1193},"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},[1188,1189,1190],{"name":1178,"slug":1179,"type":15},{"name":1164,"slug":67,"type":15},{"name":1191,"slug":1192,"type":15},"GitHub","github","2026-07-27T06:06:12.278222",{"slug":1195,"name":1195,"fn":1196,"description":1197,"org":1198,"tags":1199,"stars":1168,"repoUrl":1169,"updatedAt":1207},"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},[1200,1203,1204],{"name":1201,"slug":1202,"type":15},"Debugging","debugging",{"name":1191,"slug":1192,"type":15},{"name":1205,"slug":1206,"type":15},"Triage","triage","2026-07-14T05:25:57.442089",{"slug":1209,"name":1209,"fn":1210,"description":1211,"org":1212,"tags":1213,"stars":1168,"repoUrl":1169,"updatedAt":1220},"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},[1214,1217],{"name":1215,"slug":1216,"type":15},"Best Practices","best-practices",{"name":1218,"slug":1219,"type":15},"Code Analysis","code-analysis","2026-07-14T05:25:56.18433",{"slug":1222,"name":1222,"fn":1223,"description":1224,"org":1225,"tags":1226,"stars":1168,"repoUrl":1169,"updatedAt":1232},"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},[1227,1228,1231],{"name":13,"slug":14,"type":15},{"name":1229,"slug":1230,"type":15},"Migration","migration",{"name":9,"slug":8,"type":15},"2026-07-17T06:07:11.777011",{"slug":1234,"name":1234,"fn":1235,"description":1236,"org":1237,"tags":1238,"stars":1168,"repoUrl":1169,"updatedAt":1245},"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},[1239,1242],{"name":1240,"slug":1241,"type":15},"QA","qa",{"name":1243,"slug":1244,"type":15},"Testing","testing","2026-07-14T05:25:53.673039",{"slug":1247,"name":1247,"fn":1248,"description":1249,"org":1250,"tags":1251,"stars":1168,"repoUrl":1169,"updatedAt":1256},"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},[1252,1253],{"name":1164,"slug":67,"type":15},{"name":1254,"slug":1255,"type":15},"Infrastructure","infrastructure","2026-07-14T05:25:49.362534",{"slug":1258,"name":1258,"fn":1259,"description":1260,"org":1261,"tags":1262,"stars":1168,"repoUrl":1169,"updatedAt":1270},"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},[1263,1266,1267],{"name":1264,"slug":1265,"type":15},"Code Review","code-review",{"name":1191,"slug":1192,"type":15},{"name":1268,"slug":1269,"type":15},"Pull Requests","pull-requests","2026-07-14T05:26:01.226578",{"slug":1272,"name":1272,"fn":1273,"description":1274,"org":1275,"tags":1276,"stars":1168,"repoUrl":1169,"updatedAt":1279},"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},[1277,1278],{"name":1240,"slug":1241,"type":15},{"name":1243,"slug":1244,"type":15},"2026-07-14T05:25:54.928983",{"slug":1281,"name":1281,"fn":1282,"description":1283,"org":1284,"tags":1285,"stars":1168,"repoUrl":1169,"updatedAt":1290},"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},[1286,1289],{"name":1287,"slug":1288,"type":15},"Automation","automation",{"name":1178,"slug":1179,"type":15},"2026-07-30T05:29:03.275638",496,{"items":1293,"total":1389},[1294,1311,1321,1335,1345,1360,1375],{"slug":1295,"name":1295,"fn":1296,"description":1297,"org":1298,"tags":1299,"stars":23,"repoUrl":24,"updatedAt":1310},"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},[1300,1303,1306,1307],{"name":1301,"slug":1302,"type":15},"Data Analysis","data-analysis",{"name":1304,"slug":1305,"type":15},"Data Engineering","data-engineering",{"name":9,"slug":8,"type":15},{"name":1308,"slug":1309,"type":15},"Performance","performance","2026-07-14T05:28:43.176466",{"slug":1312,"name":1312,"fn":1313,"description":1314,"org":1315,"tags":1316,"stars":23,"repoUrl":24,"updatedAt":1320},"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},[1317,1318,1319],{"name":1164,"slug":67,"type":15},{"name":1254,"slug":1255,"type":15},{"name":9,"slug":8,"type":15},"2026-07-14T05:29:06.667109",{"slug":1322,"name":1322,"fn":1323,"description":1324,"org":1325,"tags":1326,"stars":23,"repoUrl":24,"updatedAt":1334},"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},[1327,1330,1331],{"name":1328,"slug":1329,"type":15},"Agents","agents",{"name":9,"slug":8,"type":15},{"name":1332,"slug":1333,"type":15},"Research","research","2026-07-14T05:28:06.816956",{"slug":1336,"name":1336,"fn":1337,"description":1338,"org":1339,"tags":1340,"stars":23,"repoUrl":24,"updatedAt":1344},"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},[1341,1342,1343],{"name":1301,"slug":1302,"type":15},{"name":9,"slug":8,"type":15},{"name":1243,"slug":1244,"type":15},"2026-07-17T05:29:03.913266",{"slug":1346,"name":1346,"fn":1347,"description":1348,"org":1349,"tags":1350,"stars":23,"repoUrl":24,"updatedAt":1359},"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},[1351,1352,1355,1356],{"name":1287,"slug":1288,"type":15},{"name":1353,"slug":1354,"type":15},"Imaging","imaging",{"name":9,"slug":8,"type":15},{"name":1357,"slug":1358,"type":15},"Video","video","2026-07-17T05:28:53.905004",{"slug":1361,"name":1361,"fn":1362,"description":1363,"org":1364,"tags":1365,"stars":23,"repoUrl":24,"updatedAt":1374},"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},[1366,1367,1370,1371],{"name":1164,"slug":67,"type":15},{"name":1368,"slug":1369,"type":15},"Docker","docker",{"name":9,"slug":8,"type":15},{"name":1372,"slug":1373,"type":15},"Operations","operations","2026-07-17T05:28:56.913999",{"slug":1376,"name":1376,"fn":1377,"description":1378,"org":1379,"tags":1380,"stars":23,"repoUrl":24,"updatedAt":1388},"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},[1381,1382,1385],{"name":9,"slug":8,"type":15},{"name":1383,"slug":1384,"type":15},"Quantum Computing","quantum-computing",{"name":1386,"slug":1387,"type":15},"Simulation","simulation","2026-07-14T05:26:58.898253",305]