[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill-nvidia-nemo-clinical-data-designer":3,"mdc--8na5a0-key":34,"related-repo-nvidia-nemo-clinical-data-designer":791,"related-org-nvidia-nemo-clinical-data-designer":859},{"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},"nemo-clinical-data-designer","generate synthetic clinical datasets","Use when generating synthetic tabular datasets via Data Designer — sampler columns, LLM columns, custom generators. Not for ASR audio.",{"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],{"name":13,"slug":14,"type":15},"Datasets","datasets","tag",{"name":17,"slug":18,"type":15},"Data Analysis","data-analysis",{"name":20,"slug":21,"type":15},"Life Sciences","life-sciences",{"name":9,"slug":8,"type":15},6,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002Fdigital-health-skills","2026-07-14T05:35:42.770246","Apache-2.0",1,[],{"repoUrl":24,"stars":23,"forks":27,"topics":30,"description":31},[],"NVIDIA Digital Health skills repo for agent-guided healthcare AI workflows, including evaluation, data generation, model adaptation, deployment guidance, and developer best practices.","https:\u002F\u002Fgithub.com\u002FNVIDIA\u002Fdigital-health-skills\u002Ftree\u002FHEAD\u002Fskills\u002Fnemo-clinical-data-designer","---\nname: nemo-clinical-data-designer\nlicense: \"Apache-2.0\"\ndescription: \"Use when generating synthetic tabular datasets via Data Designer — sampler columns, LLM columns, custom generators. Not for ASR audio.\"\nargument-hint: [describe the dataset you want to generate]\nmetadata:\n  author: \"Voice Eval Flywheel team \u003Cbrandoing@nvidia.com>\"\n  tags:\n    - nvidia\n    - data-designer\n    - synthetic-data\n    - dataset-generation\n    - llm\n    - sampler\n---\n\n# NeMo Clinical Data Designer\n\n## Purpose\n\nProduce a synthetic tabular dataset by authoring a NeMo Data Designer config (`data_designer.config`), previewing the output, then committing to a full run. The agent first picks a conversation mode (Interactive vs. Autopilot), drives the user through config construction, and ends at a written-to-disk dataset.\n\n## Instructions\n\nThe flow is five steps, in order:\n\n1. Decide on **conversation mode**. Default is Interactive — ask the user clarifying questions as you build. Switch to Autopilot whenever the user signals they want an opinionated build with minimal back-and-forth (\"just build it\", \"you decide\", \"surprise me\").\n2. Open the matching per-mode runbook — `workflows\u002Finteractive.md` or `workflows\u002Fautopilot.md` — and follow it. The runbooks are the source of truth for which questions to ask and what defaults to pick.\n3. Author a Python config module exposing `load_config_builder()` that returns a `DataDesignerConfigBuilder`. The Output Template section near the end of this file shows the skeleton (PEP 723 deps header + samplers + LLM columns + optional Pydantic schema + custom generators).\n4. Preview before generating: run `data-designer preview` against the config and show the user the sample rows.\n5. After approval, run `data-designer run` (the `run_script` entry point) for the full generation. If the dataset references person attributes, first inspect the person-object schema via `python3 scripts\u002Fget_person_object_schema.py`.\n\n## Examples\n\nWorked example dialogues live in the per-mode runbooks; the SKILL.md doesn't duplicate them:\n\n- For a question-driven build with full user collaboration → `workflows\u002Finteractive.md`.\n- For a hands-off \"you decide\" build → `workflows\u002Fautopilot.md`.\n\nA representative Python config — sampler column + LLM column + Pydantic-typed output + custom column generator — is laid out in the Output Template section further down.\n\n## Scripts\n\n| Script | Purpose |\n|---|---|\n| `scripts\u002Fget_person_object_schema.py` | Inspect the Data Designer person-object field schema (names, demographics, addresses) before adding person-sampler columns to a config. Invoke via `python3 scripts\u002Fget_person_object_schema.py`. |\n\n## Prerequisites\n\n- A Python ≥ 3.10 virtualenv (Data Designer requires it).\n- The `data-designer` package itself — install via `pip install data-designer`.\n- If the config you build references LLM columns or LLM judges, the user must have valid LLM credentials configured per Data Designer's own auth conventions.\n- `pydantic` is only needed when the config declares structured outputs; the PEP 723 header in the generated script will fetch it automatically.\n\n## Limitations\n\n- Output domain is tabular synthetic data exclusively — no audio, images, or unstructured documents. Clinical-ASR audio benchmarks belong to `\u002Fdigital-health-clinical-asr-build`, not here.\n- The skill writes the Python config file but does not invoke the generation pipeline on the user's behalf — the user runs `data-designer preview` and `data-designer run` themselves.\n- LLM columns require network reachability at preview and run time. In sandboxed environments those calls may be blocked; see Troubleshooting for the sandbox-disable handshake.\n\n## Brief to satisfy\n\n$ARGUMENTS\n\nThe brief above is what you need to translate into a Data Designer config. Don't open additional workspace files to \"look around\" — the selected workflow runbook walks you through any context-gathering you actually need (its Learn step). Trust the runbook over instinct here.\n\n## Mode selection + runbook\n\nA simple heuristic: if the user's framing is opinionated-and-hands-off (\"be opinionated\", \"you decide\", \"make reasonable assumptions\", \"just build it\", \"surprise me\", and similar), pick **Autopilot**. For everything else, default to **Interactive** so the user shapes column choices.\n\nAfter selecting, read *only* that mode's runbook and execute it linearly:\n\n- Interactive path → `workflows\u002Finteractive.md`.\n- Autopilot path → `workflows\u002Fautopilot.md`.\n\n## Authoring rules\n\nA handful of constraints that hold regardless of which mode you're in:\n\n- Output column retention defaults to *keep everything*. Only drop a column when the user explicitly asks, or when the column is a pure intermediate (e.g., a person-object sampler whose fields you extract into discrete columns downstream). Err toward keeping; the user can always trim later.\n- Seed datasets are never suggested unilaterally. Only wire one in when the user volunteers seed data or asks to build off existing records. If they do, consult `references\u002Fseed-datasets.md` before configuring the seed source.\n- Person attributes (names, demographics, addresses) require the person-sampling reference — read `references\u002Fperson-sampling.md` before adding any person columns.\n- If a config script already in the workspace matches the brief, surface the conflict to the user and ask whether to edit it in place or author a new one. Never silently overwrite.\n\n## Usage Tips and Common Pitfalls\n\n- **Sampler \u002F validation columns require both a type and a params object.** Pair `sampler_type=\"category\"` with `params=dd.CategorySamplerParams(...)`; either alone won't validate.\n- **Jinja2 in `prompt`, `system_prompt`, and `expr`** — top-level columns are `{{ column_name }}`, nested fields are `{{ column_name.field }}`.\n- **`SamplerColumnConfig` constructor takes `params`** — not `sampler_params`. Easy to typo if you're new to the library.\n- **Reading LLM-judge scores.** `LLMJudgeColumnConfig` returns a nested dict shaped `{score_name: {reasoning: str, score: int}}`. To pull the integer score in a Jinja template for a judge column `quality` with score `correctness`, write `{{ quality.correctness.score }}`. Dropping the trailing `.score` yields the full dict, which is almost never what you want.\n\n## Troubleshooting\n\n- **`data-designer: command not found`** → the package isn't installed in this environment. Tell the user, confirm they're on Python ≥ 3.10, and offer to create a venv + run `pip install data-designer` for them — but wait for explicit permission before installing.\n- **Preview fails on network errors** → most likely a sandbox is blocking outbound calls to the LLM endpoint. Ask whether to retry with the sandbox disabled. If that still fails, ask the user to run the command in their own shell.\n\n## Output deliverable\n\nA single Python file in the user's working directory exporting `load_config_builder()` → `DataDesignerConfigBuilder`. Filename should describe the dataset (`customer_reviews.py`, `medical_term_seeds.py`), not be generic. Use a PEP 723 inline metadata dep header so the file is runnable with `data-designer run` directly.\n\nFor the canonical skeleton — Pydantic-typed output, custom column generators, optional seed-dataset wiring — read `references\u002Foutput-template.md`. Strip blocks from the skeleton when the brief doesn't need them.\n\n",{"data":35,"body":46},{"name":4,"license":26,"description":6,"argument-hint":36,"metadata":38},[37],"describe the dataset you want to generate",{"author":39,"tags":40},"Voice Eval Flywheel team \u003Cbrandoing@nvidia.com>",[8,41,42,43,44,45],"data-designer","synthetic-data","dataset-generation","llm","sampler",{"type":47,"children":48},"root",[49,57,64,79,85,90,192,198,203,229,234,240,290,296,339,345,385,391,396,401,407,426,439,464,470,475,521,527,691,697,731,737,778],{"type":50,"tag":51,"props":52,"children":53},"element","h1",{"id":4},[54],{"type":55,"value":56},"text","NeMo Clinical Data Designer",{"type":50,"tag":58,"props":59,"children":61},"h2",{"id":60},"purpose",[62],{"type":55,"value":63},"Purpose",{"type":50,"tag":65,"props":66,"children":67},"p",{},[68,70,77],{"type":55,"value":69},"Produce a synthetic tabular dataset by authoring a NeMo Data Designer config (",{"type":50,"tag":71,"props":72,"children":74},"code",{"className":73},[],[75],{"type":55,"value":76},"data_designer.config",{"type":55,"value":78},"), previewing the output, then committing to a full run. The agent first picks a conversation mode (Interactive vs. Autopilot), drives the user through config construction, and ends at a written-to-disk dataset.",{"type":50,"tag":58,"props":80,"children":82},{"id":81},"instructions",[83],{"type":55,"value":84},"Instructions",{"type":50,"tag":65,"props":86,"children":87},{},[88],{"type":55,"value":89},"The flow is five steps, in order:",{"type":50,"tag":91,"props":92,"children":93},"ol",{},[94,108,129,150,163],{"type":50,"tag":95,"props":96,"children":97},"li",{},[98,100,106],{"type":55,"value":99},"Decide on ",{"type":50,"tag":101,"props":102,"children":103},"strong",{},[104],{"type":55,"value":105},"conversation mode",{"type":55,"value":107},". Default is Interactive — ask the user clarifying questions as you build. Switch to Autopilot whenever the user signals they want an opinionated build with minimal back-and-forth (\"just build it\", \"you decide\", \"surprise me\").",{"type":50,"tag":95,"props":109,"children":110},{},[111,113,119,121,127],{"type":55,"value":112},"Open the matching per-mode runbook — ",{"type":50,"tag":71,"props":114,"children":116},{"className":115},[],[117],{"type":55,"value":118},"workflows\u002Finteractive.md",{"type":55,"value":120}," or ",{"type":50,"tag":71,"props":122,"children":124},{"className":123},[],[125],{"type":55,"value":126},"workflows\u002Fautopilot.md",{"type":55,"value":128}," — and follow it. The runbooks are the source of truth for which questions to ask and what defaults to pick.",{"type":50,"tag":95,"props":130,"children":131},{},[132,134,140,142,148],{"type":55,"value":133},"Author a Python config module exposing ",{"type":50,"tag":71,"props":135,"children":137},{"className":136},[],[138],{"type":55,"value":139},"load_config_builder()",{"type":55,"value":141}," that returns a ",{"type":50,"tag":71,"props":143,"children":145},{"className":144},[],[146],{"type":55,"value":147},"DataDesignerConfigBuilder",{"type":55,"value":149},". The Output Template section near the end of this file shows the skeleton (PEP 723 deps header + samplers + LLM columns + optional Pydantic schema + custom generators).",{"type":50,"tag":95,"props":151,"children":152},{},[153,155,161],{"type":55,"value":154},"Preview before generating: run ",{"type":50,"tag":71,"props":156,"children":158},{"className":157},[],[159],{"type":55,"value":160},"data-designer preview",{"type":55,"value":162}," against the config and show the user the sample rows.",{"type":50,"tag":95,"props":164,"children":165},{},[166,168,174,176,182,184,190],{"type":55,"value":167},"After approval, run ",{"type":50,"tag":71,"props":169,"children":171},{"className":170},[],[172],{"type":55,"value":173},"data-designer run",{"type":55,"value":175}," (the ",{"type":50,"tag":71,"props":177,"children":179},{"className":178},[],[180],{"type":55,"value":181},"run_script",{"type":55,"value":183}," entry point) for the full generation. If the dataset references person attributes, first inspect the person-object schema via ",{"type":50,"tag":71,"props":185,"children":187},{"className":186},[],[188],{"type":55,"value":189},"python3 scripts\u002Fget_person_object_schema.py",{"type":55,"value":191},".",{"type":50,"tag":58,"props":193,"children":195},{"id":194},"examples",[196],{"type":55,"value":197},"Examples",{"type":50,"tag":65,"props":199,"children":200},{},[201],{"type":55,"value":202},"Worked example dialogues live in the per-mode runbooks; the SKILL.md doesn't duplicate them:",{"type":50,"tag":204,"props":205,"children":206},"ul",{},[207,218],{"type":50,"tag":95,"props":208,"children":209},{},[210,212,217],{"type":55,"value":211},"For a question-driven build with full user collaboration → ",{"type":50,"tag":71,"props":213,"children":215},{"className":214},[],[216],{"type":55,"value":118},{"type":55,"value":191},{"type":50,"tag":95,"props":219,"children":220},{},[221,223,228],{"type":55,"value":222},"For a hands-off \"you decide\" build → ",{"type":50,"tag":71,"props":224,"children":226},{"className":225},[],[227],{"type":55,"value":126},{"type":55,"value":191},{"type":50,"tag":65,"props":230,"children":231},{},[232],{"type":55,"value":233},"A representative Python config — sampler column + LLM column + Pydantic-typed output + custom column generator — is laid out in the Output Template section further down.",{"type":50,"tag":58,"props":235,"children":237},{"id":236},"scripts",[238],{"type":55,"value":239},"Scripts",{"type":50,"tag":241,"props":242,"children":243},"table",{},[244,262],{"type":50,"tag":245,"props":246,"children":247},"thead",{},[248],{"type":50,"tag":249,"props":250,"children":251},"tr",{},[252,258],{"type":50,"tag":253,"props":254,"children":255},"th",{},[256],{"type":55,"value":257},"Script",{"type":50,"tag":253,"props":259,"children":260},{},[261],{"type":55,"value":63},{"type":50,"tag":263,"props":264,"children":265},"tbody",{},[266],{"type":50,"tag":249,"props":267,"children":268},{},[269,279],{"type":50,"tag":270,"props":271,"children":272},"td",{},[273],{"type":50,"tag":71,"props":274,"children":276},{"className":275},[],[277],{"type":55,"value":278},"scripts\u002Fget_person_object_schema.py",{"type":50,"tag":270,"props":280,"children":281},{},[282,284,289],{"type":55,"value":283},"Inspect the Data Designer person-object field schema (names, demographics, addresses) before adding person-sampler columns to a config. Invoke via ",{"type":50,"tag":71,"props":285,"children":287},{"className":286},[],[288],{"type":55,"value":189},{"type":55,"value":191},{"type":50,"tag":58,"props":291,"children":293},{"id":292},"prerequisites",[294],{"type":55,"value":295},"Prerequisites",{"type":50,"tag":204,"props":297,"children":298},{},[299,304,323,328],{"type":50,"tag":95,"props":300,"children":301},{},[302],{"type":55,"value":303},"A Python ≥ 3.10 virtualenv (Data Designer requires it).",{"type":50,"tag":95,"props":305,"children":306},{},[307,309,314,316,322],{"type":55,"value":308},"The ",{"type":50,"tag":71,"props":310,"children":312},{"className":311},[],[313],{"type":55,"value":41},{"type":55,"value":315}," package itself — install via ",{"type":50,"tag":71,"props":317,"children":319},{"className":318},[],[320],{"type":55,"value":321},"pip install data-designer",{"type":55,"value":191},{"type":50,"tag":95,"props":324,"children":325},{},[326],{"type":55,"value":327},"If the config you build references LLM columns or LLM judges, the user must have valid LLM credentials configured per Data Designer's own auth conventions.",{"type":50,"tag":95,"props":329,"children":330},{},[331,337],{"type":50,"tag":71,"props":332,"children":334},{"className":333},[],[335],{"type":55,"value":336},"pydantic",{"type":55,"value":338}," is only needed when the config declares structured outputs; the PEP 723 header in the generated script will fetch it automatically.",{"type":50,"tag":58,"props":340,"children":342},{"id":341},"limitations",[343],{"type":55,"value":344},"Limitations",{"type":50,"tag":204,"props":346,"children":347},{},[348,361,380],{"type":50,"tag":95,"props":349,"children":350},{},[351,353,359],{"type":55,"value":352},"Output domain is tabular synthetic data exclusively — no audio, images, or unstructured documents. Clinical-ASR audio benchmarks belong to ",{"type":50,"tag":71,"props":354,"children":356},{"className":355},[],[357],{"type":55,"value":358},"\u002Fdigital-health-clinical-asr-build",{"type":55,"value":360},", not here.",{"type":50,"tag":95,"props":362,"children":363},{},[364,366,371,373,378],{"type":55,"value":365},"The skill writes the Python config file but does not invoke the generation pipeline on the user's behalf — the user runs ",{"type":50,"tag":71,"props":367,"children":369},{"className":368},[],[370],{"type":55,"value":160},{"type":55,"value":372}," and ",{"type":50,"tag":71,"props":374,"children":376},{"className":375},[],[377],{"type":55,"value":173},{"type":55,"value":379}," themselves.",{"type":50,"tag":95,"props":381,"children":382},{},[383],{"type":55,"value":384},"LLM columns require network reachability at preview and run time. In sandboxed environments those calls may be blocked; see Troubleshooting for the sandbox-disable handshake.",{"type":50,"tag":58,"props":386,"children":388},{"id":387},"brief-to-satisfy",[389],{"type":55,"value":390},"Brief to satisfy",{"type":50,"tag":65,"props":392,"children":393},{},[394],{"type":55,"value":395},"$ARGUMENTS",{"type":50,"tag":65,"props":397,"children":398},{},[399],{"type":55,"value":400},"The brief above is what you need to translate into a Data Designer config. Don't open additional workspace files to \"look around\" — the selected workflow runbook walks you through any context-gathering you actually need (its Learn step). Trust the runbook over instinct here.",{"type":50,"tag":58,"props":402,"children":404},{"id":403},"mode-selection-runbook",[405],{"type":55,"value":406},"Mode selection + runbook",{"type":50,"tag":65,"props":408,"children":409},{},[410,412,417,419,424],{"type":55,"value":411},"A simple heuristic: if the user's framing is opinionated-and-hands-off (\"be opinionated\", \"you decide\", \"make reasonable assumptions\", \"just build it\", \"surprise me\", and similar), pick ",{"type":50,"tag":101,"props":413,"children":414},{},[415],{"type":55,"value":416},"Autopilot",{"type":55,"value":418},". For everything else, default to ",{"type":50,"tag":101,"props":420,"children":421},{},[422],{"type":55,"value":423},"Interactive",{"type":55,"value":425}," so the user shapes column choices.",{"type":50,"tag":65,"props":427,"children":428},{},[429,431,437],{"type":55,"value":430},"After selecting, read ",{"type":50,"tag":432,"props":433,"children":434},"em",{},[435],{"type":55,"value":436},"only",{"type":55,"value":438}," that mode's runbook and execute it linearly:",{"type":50,"tag":204,"props":440,"children":441},{},[442,453],{"type":50,"tag":95,"props":443,"children":444},{},[445,447,452],{"type":55,"value":446},"Interactive path → ",{"type":50,"tag":71,"props":448,"children":450},{"className":449},[],[451],{"type":55,"value":118},{"type":55,"value":191},{"type":50,"tag":95,"props":454,"children":455},{},[456,458,463],{"type":55,"value":457},"Autopilot path → ",{"type":50,"tag":71,"props":459,"children":461},{"className":460},[],[462],{"type":55,"value":126},{"type":55,"value":191},{"type":50,"tag":58,"props":465,"children":467},{"id":466},"authoring-rules",[468],{"type":55,"value":469},"Authoring rules",{"type":50,"tag":65,"props":471,"children":472},{},[473],{"type":55,"value":474},"A handful of constraints that hold regardless of which mode you're in:",{"type":50,"tag":204,"props":476,"children":477},{},[478,490,503,516],{"type":50,"tag":95,"props":479,"children":480},{},[481,483,488],{"type":55,"value":482},"Output column retention defaults to ",{"type":50,"tag":432,"props":484,"children":485},{},[486],{"type":55,"value":487},"keep everything",{"type":55,"value":489},". Only drop a column when the user explicitly asks, or when the column is a pure intermediate (e.g., a person-object sampler whose fields you extract into discrete columns downstream). Err toward keeping; the user can always trim later.",{"type":50,"tag":95,"props":491,"children":492},{},[493,495,501],{"type":55,"value":494},"Seed datasets are never suggested unilaterally. Only wire one in when the user volunteers seed data or asks to build off existing records. If they do, consult ",{"type":50,"tag":71,"props":496,"children":498},{"className":497},[],[499],{"type":55,"value":500},"references\u002Fseed-datasets.md",{"type":55,"value":502}," before configuring the seed source.",{"type":50,"tag":95,"props":504,"children":505},{},[506,508,514],{"type":55,"value":507},"Person attributes (names, demographics, addresses) require the person-sampling reference — read ",{"type":50,"tag":71,"props":509,"children":511},{"className":510},[],[512],{"type":55,"value":513},"references\u002Fperson-sampling.md",{"type":55,"value":515}," before adding any person columns.",{"type":50,"tag":95,"props":517,"children":518},{},[519],{"type":55,"value":520},"If a config script already in the workspace matches the brief, surface the conflict to the user and ask whether to edit it in place or author a new one. Never silently overwrite.",{"type":50,"tag":58,"props":522,"children":524},{"id":523},"usage-tips-and-common-pitfalls",[525],{"type":55,"value":526},"Usage Tips and Common Pitfalls",{"type":50,"tag":204,"props":528,"children":529},{},[530,556,603,633],{"type":50,"tag":95,"props":531,"children":532},{},[533,538,540,546,548,554],{"type":50,"tag":101,"props":534,"children":535},{},[536],{"type":55,"value":537},"Sampler \u002F validation columns require both a type and a params object.",{"type":55,"value":539}," Pair ",{"type":50,"tag":71,"props":541,"children":543},{"className":542},[],[544],{"type":55,"value":545},"sampler_type=\"category\"",{"type":55,"value":547}," with ",{"type":50,"tag":71,"props":549,"children":551},{"className":550},[],[552],{"type":55,"value":553},"params=dd.CategorySamplerParams(...)",{"type":55,"value":555},"; either alone won't validate.",{"type":50,"tag":95,"props":557,"children":558},{},[559,586,588,594,596,602],{"type":50,"tag":101,"props":560,"children":561},{},[562,564,570,572,578,580],{"type":55,"value":563},"Jinja2 in ",{"type":50,"tag":71,"props":565,"children":567},{"className":566},[],[568],{"type":55,"value":569},"prompt",{"type":55,"value":571},", ",{"type":50,"tag":71,"props":573,"children":575},{"className":574},[],[576],{"type":55,"value":577},"system_prompt",{"type":55,"value":579},", and ",{"type":50,"tag":71,"props":581,"children":583},{"className":582},[],[584],{"type":55,"value":585},"expr",{"type":55,"value":587}," — top-level columns are ",{"type":50,"tag":71,"props":589,"children":591},{"className":590},[],[592],{"type":55,"value":593},"{{ column_name }}",{"type":55,"value":595},", nested fields are ",{"type":50,"tag":71,"props":597,"children":599},{"className":598},[],[600],{"type":55,"value":601},"{{ column_name.field }}",{"type":55,"value":191},{"type":50,"tag":95,"props":604,"children":605},{},[606,623,625,631],{"type":50,"tag":101,"props":607,"children":608},{},[609,615,617],{"type":50,"tag":71,"props":610,"children":612},{"className":611},[],[613],{"type":55,"value":614},"SamplerColumnConfig",{"type":55,"value":616}," constructor takes ",{"type":50,"tag":71,"props":618,"children":620},{"className":619},[],[621],{"type":55,"value":622},"params",{"type":55,"value":624}," — not ",{"type":50,"tag":71,"props":626,"children":628},{"className":627},[],[629],{"type":55,"value":630},"sampler_params",{"type":55,"value":632},". Easy to typo if you're new to the library.",{"type":50,"tag":95,"props":634,"children":635},{},[636,641,643,649,651,657,659,665,667,673,675,681,683,689],{"type":50,"tag":101,"props":637,"children":638},{},[639],{"type":55,"value":640},"Reading LLM-judge scores.",{"type":55,"value":642}," ",{"type":50,"tag":71,"props":644,"children":646},{"className":645},[],[647],{"type":55,"value":648},"LLMJudgeColumnConfig",{"type":55,"value":650}," returns a nested dict shaped ",{"type":50,"tag":71,"props":652,"children":654},{"className":653},[],[655],{"type":55,"value":656},"{score_name: {reasoning: str, score: int}}",{"type":55,"value":658},". To pull the integer score in a Jinja template for a judge column ",{"type":50,"tag":71,"props":660,"children":662},{"className":661},[],[663],{"type":55,"value":664},"quality",{"type":55,"value":666}," with score ",{"type":50,"tag":71,"props":668,"children":670},{"className":669},[],[671],{"type":55,"value":672},"correctness",{"type":55,"value":674},", write ",{"type":50,"tag":71,"props":676,"children":678},{"className":677},[],[679],{"type":55,"value":680},"{{ quality.correctness.score }}",{"type":55,"value":682},". Dropping the trailing ",{"type":50,"tag":71,"props":684,"children":686},{"className":685},[],[687],{"type":55,"value":688},".score",{"type":55,"value":690}," yields the full dict, which is almost never what you want.",{"type":50,"tag":58,"props":692,"children":694},{"id":693},"troubleshooting",[695],{"type":55,"value":696},"Troubleshooting",{"type":50,"tag":204,"props":698,"children":699},{},[700,721],{"type":50,"tag":95,"props":701,"children":702},{},[703,712,714,719],{"type":50,"tag":101,"props":704,"children":705},{},[706],{"type":50,"tag":71,"props":707,"children":709},{"className":708},[],[710],{"type":55,"value":711},"data-designer: command not found",{"type":55,"value":713}," → the package isn't installed in this environment. Tell the user, confirm they're on Python ≥ 3.10, and offer to create a venv + run ",{"type":50,"tag":71,"props":715,"children":717},{"className":716},[],[718],{"type":55,"value":321},{"type":55,"value":720}," for them — but wait for explicit permission before installing.",{"type":50,"tag":95,"props":722,"children":723},{},[724,729],{"type":50,"tag":101,"props":725,"children":726},{},[727],{"type":55,"value":728},"Preview fails on network errors",{"type":55,"value":730}," → most likely a sandbox is blocking outbound calls to the LLM endpoint. Ask whether to retry with the sandbox disabled. If that still fails, ask the user to run the command in their own shell.",{"type":50,"tag":58,"props":732,"children":734},{"id":733},"output-deliverable",[735],{"type":55,"value":736},"Output deliverable",{"type":50,"tag":65,"props":738,"children":739},{},[740,742,747,749,754,756,762,763,769,771,776],{"type":55,"value":741},"A single Python file in the user's working directory exporting ",{"type":50,"tag":71,"props":743,"children":745},{"className":744},[],[746],{"type":55,"value":139},{"type":55,"value":748}," → ",{"type":50,"tag":71,"props":750,"children":752},{"className":751},[],[753],{"type":55,"value":147},{"type":55,"value":755},". Filename should describe the dataset (",{"type":50,"tag":71,"props":757,"children":759},{"className":758},[],[760],{"type":55,"value":761},"customer_reviews.py",{"type":55,"value":571},{"type":50,"tag":71,"props":764,"children":766},{"className":765},[],[767],{"type":55,"value":768},"medical_term_seeds.py",{"type":55,"value":770},"), not be generic. Use a PEP 723 inline metadata dep header so the file is runnable with ",{"type":50,"tag":71,"props":772,"children":774},{"className":773},[],[775],{"type":55,"value":173},{"type":55,"value":777}," directly.",{"type":50,"tag":65,"props":779,"children":780},{},[781,783,789],{"type":55,"value":782},"For the canonical skeleton — Pydantic-typed output, custom column generators, optional seed-dataset wiring — read ",{"type":50,"tag":71,"props":784,"children":786},{"className":785},[],[787],{"type":55,"value":788},"references\u002Foutput-template.md",{"type":55,"value":790},". Strip blocks from the skeleton when the brief doesn't need them.",{"items":792,"total":858},[793,800,817,830,843],{"slug":4,"name":4,"fn":5,"description":6,"org":794,"tags":795,"stars":23,"repoUrl":24,"updatedAt":25},{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[796,797,798,799],{"name":17,"slug":18,"type":15},{"name":13,"slug":14,"type":15},{"name":20,"slug":21,"type":15},{"name":9,"slug":8,"type":15},{"slug":801,"name":801,"fn":802,"description":803,"org":804,"tags":805,"stars":23,"repoUrl":24,"updatedAt":816},"riva-asr","run and test Riva ASR models","Use when the user wants to deploy, run, or test an ASR (speech-to-text) Riva NIM — cloud-hosted (build.nvidia.com) or self-hosted Parakeet\u002FCanary\u002FWhisper.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[806,809,810,813],{"name":807,"slug":808,"type":15},"AI Infrastructure","ai-infrastructure",{"name":9,"slug":8,"type":15},{"name":811,"slug":812,"type":15},"Speech","speech",{"name":814,"slug":815,"type":15},"Testing","testing","2026-07-14T05:35:49.042673",{"slug":818,"name":818,"fn":819,"description":820,"org":821,"tags":822,"stars":23,"repoUrl":24,"updatedAt":829},"riva-asr-custom","deploy custom Riva ASR models","Use when the user wants to deploy a custom-trained ASR model as a Riva NIM, or convert a NeMo model via nemo2riva \u002F riva-build \u002F riva-deploy \u002F RMIR.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[823,824,827,828],{"name":807,"slug":808,"type":15},{"name":825,"slug":826,"type":15},"Deployment","deployment",{"name":9,"slug":8,"type":15},{"name":811,"slug":812,"type":15},"2026-07-14T05:35:44.009948",{"slug":831,"name":831,"fn":832,"description":833,"org":834,"tags":835,"stars":23,"repoUrl":24,"updatedAt":842},"riva-nim-setup","set up NVIDIA Riva Speech NIMs","Use when getting started with NVIDIA Riva Speech NIMs: NGC access, Docker login for nvcr.io, NVIDIA Container Toolkit, GPU verification, Riva Python client.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[836,837,840,841],{"name":807,"slug":808,"type":15},{"name":838,"slug":839,"type":15},"Docker","docker",{"name":9,"slug":8,"type":15},{"name":811,"slug":812,"type":15},"2026-07-14T05:35:50.295877",{"slug":844,"name":844,"fn":845,"description":846,"org":847,"tags":848,"stars":23,"repoUrl":24,"updatedAt":857},"riva-tts","deploy and run Riva TTS models","Use when the user wants to deploy, run, or test a TTS (speech-synthesis) Riva NIM — cloud-hosted (build.nvidia.com) or self-hosted Magpie \u002F voice cloning.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},[849,850,853,854],{"name":807,"slug":808,"type":15},{"name":851,"slug":852,"type":15},"Audio","audio",{"name":9,"slug":8,"type":15},{"name":855,"slug":856,"type":15},"Text-to-Speech","text-to-speech","2026-07-14T05:35:45.262923",5,{"items":860,"total":1015},[861,879,895,906,918,932,945,959,970,981,995,1004],{"slug":862,"name":862,"fn":863,"description":864,"org":865,"tags":866,"stars":876,"repoUrl":877,"updatedAt":878},"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},[867,870,873],{"name":868,"slug":869,"type":15},"Documentation","documentation",{"name":871,"slug":872,"type":15},"MCP","mcp",{"name":874,"slug":875,"type":15},"Search","search",21777,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FNemoClaw","2026-07-20T06:00:01.461044",{"slug":880,"name":880,"fn":881,"description":882,"org":883,"tags":884,"stars":892,"repoUrl":893,"updatedAt":894},"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},[885,888,889],{"name":886,"slug":887,"type":15},"Containers","containers",{"name":825,"slug":826,"type":15},{"name":890,"slug":891,"type":15},"Python","python",17049,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FMegatron-LM","2026-07-27T06:06:11.249662",{"slug":896,"name":896,"fn":897,"description":898,"org":899,"tags":900,"stars":892,"repoUrl":893,"updatedAt":905},"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},[901,904],{"name":902,"slug":903,"type":15},"CI\u002FCD","ci-cd",{"name":825,"slug":826,"type":15},"2026-07-14T05:25:59.97109",{"slug":907,"name":907,"fn":908,"description":909,"org":910,"tags":911,"stars":892,"repoUrl":893,"updatedAt":917},"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},[912,913,914],{"name":902,"slug":903,"type":15},{"name":825,"slug":826,"type":15},{"name":915,"slug":916,"type":15},"GitHub","github","2026-07-27T06:06:12.278222",{"slug":919,"name":919,"fn":920,"description":921,"org":922,"tags":923,"stars":892,"repoUrl":893,"updatedAt":931},"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},[924,927,928],{"name":925,"slug":926,"type":15},"Debugging","debugging",{"name":915,"slug":916,"type":15},{"name":929,"slug":930,"type":15},"Triage","triage","2026-07-14T05:25:57.442089",{"slug":933,"name":933,"fn":934,"description":935,"org":936,"tags":937,"stars":892,"repoUrl":893,"updatedAt":944},"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},[938,941],{"name":939,"slug":940,"type":15},"Best Practices","best-practices",{"name":942,"slug":943,"type":15},"Code Analysis","code-analysis","2026-07-14T05:25:56.18433",{"slug":946,"name":946,"fn":947,"description":948,"org":949,"tags":950,"stars":892,"repoUrl":893,"updatedAt":958},"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},[951,954,957],{"name":952,"slug":953,"type":15},"Machine Learning","machine-learning",{"name":955,"slug":956,"type":15},"Migration","migration",{"name":9,"slug":8,"type":15},"2026-07-17T06:07:11.777011",{"slug":960,"name":960,"fn":961,"description":962,"org":963,"tags":964,"stars":892,"repoUrl":893,"updatedAt":969},"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},[965,968],{"name":966,"slug":967,"type":15},"QA","qa",{"name":814,"slug":815,"type":15},"2026-07-14T05:25:53.673039",{"slug":971,"name":971,"fn":972,"description":973,"org":974,"tags":975,"stars":892,"repoUrl":893,"updatedAt":980},"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},[976,977],{"name":825,"slug":826,"type":15},{"name":978,"slug":979,"type":15},"Infrastructure","infrastructure","2026-07-14T05:25:49.362534",{"slug":982,"name":982,"fn":983,"description":984,"org":985,"tags":986,"stars":892,"repoUrl":893,"updatedAt":994},"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},[987,990,991],{"name":988,"slug":989,"type":15},"Code Review","code-review",{"name":915,"slug":916,"type":15},{"name":992,"slug":993,"type":15},"Pull Requests","pull-requests","2026-07-14T05:26:01.226578",{"slug":996,"name":996,"fn":997,"description":998,"org":999,"tags":1000,"stars":892,"repoUrl":893,"updatedAt":1003},"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},[1001,1002],{"name":966,"slug":967,"type":15},{"name":814,"slug":815,"type":15},"2026-07-14T05:25:54.928983",{"slug":1005,"name":1005,"fn":1006,"description":1007,"org":1008,"tags":1009,"stars":892,"repoUrl":893,"updatedAt":1014},"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},[1010,1013],{"name":1011,"slug":1012,"type":15},"Automation","automation",{"name":902,"slug":903,"type":15},"2026-07-30T05:29:03.275638",496]