[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill-nvidia-tao-generate-referring-expressions":3,"mdc-mtx8a0-key":31,"related-org-nvidia-tao-generate-referring-expressions":1698,"related-repo-nvidia-tao-generate-referring-expressions":1859},{"slug":4,"name":4,"fn":5,"description":6,"org":7,"tags":11,"stars":20,"repoUrl":21,"updatedAt":22,"license":23,"forks":24,"topics":25,"repo":26,"sourceUrl":29,"mdContent":30},"tao-generate-referring-expressions","generate referring expressions for images","Four-step image referring-expression pipeline: turns images plus KITTI bounding-box labels into region descriptions, scene captions, grounded referring expressions, and (optionally) verified expressions via VLM distillation. Use when the user wants to generate referring-expression annotations from images with KITTI labels, build region descriptions, produce grouped grounding phrases tied to bboxes, run a double-check verification pass on grounding expressions, auto-label traffic \u002F scene images for referring datasets, or run the image_referring_expression pipeline. Triggers include 'referring expression', 'region description', 'KITTI labels', 'spatial relationship annotation', 'auto-label image referring expression', 'image_referring_expression'.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":9},"nvidia","NVIDIA","https:\u002F\u002Fpexgzepcugksgbtrxkhf.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Forg-logos\u002Fnvidia.png",[12,16,17],{"name":13,"slug":14,"type":15},"Deep Learning","deep-learning","tag",{"name":9,"slug":8,"type":15},{"name":18,"slug":19,"type":15},"Computer Vision","computer-vision",2473,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002Fskills","2026-07-14T05:30:30.434263","Apache-2.0",281,[],{"repoUrl":21,"stars":20,"forks":24,"topics":27,"description":28},[],"AI agent skills published by NVIDIA","https:\u002F\u002Fgithub.com\u002FNVIDIA\u002Fskills\u002Ftree\u002FHEAD\u002Fskills\u002Ftao-generate-referring-expressions","---\nname: tao-generate-referring-expressions\ndescription: \"Four-step image referring-expression pipeline: turns images plus KITTI bounding-box labels into region\n  descriptions, scene captions, grounded referring expressions, and (optionally) verified expressions via VLM distillation. Use\n  when the user wants to generate referring-expression annotations from images with KITTI labels, build region descriptions,\n  produce grouped grounding phrases tied to bboxes, run a double-check verification pass on grounding expressions, auto-label\n  traffic \u002F scene images for referring datasets, or run the image_referring_expression pipeline. Triggers include 'referring\n  expression', 'region description', 'KITTI labels', 'spatial relationship annotation', 'auto-label image referring expression',\n  'image_referring_expression'.\"\nlicense: Apache-2.0\ncompatibility: Requires docker + nvidia-container-toolkit + at least one VLM endpoint (Gemini API key or OpenAI-compatible).\nmetadata:\n  author: NVIDIA Corporation\n  version: \"0.1.0\"\ntags:\n  - image\n  - referring-expression\n  - kitti\n  - bounding-boxes\n  - auto-label\n  - vlm\nallowed-tools: Read Bash Write\n---\n\n# Image Referring Expression Pipeline\n\nGenerate referring-expression and grounding annotations from images with KITTI-format bounding box labels. A single VLM (Gemini or any OpenAI-compatible endpoint) runs four steps: per-object region descriptions, holistic image captions, grouped grounding expressions tied to bboxes, and an optional double-check verification pass.\n\n## Purpose\n\nTransform `(image, KITTI labels)` pairs into a unified `annotations.jsonl` containing rich, grounded referring expressions. The VLM acts as a \"teacher\" annotator: Steps 0-1 see the image; Step 2 groups Step 0 outputs into grouping phrases with bbox lists; Step 3 (optional) re-examines those bboxes against the image and corrects mismatches.\n\n## Pipeline Architecture\n\n```\nStep 0: Region expression  ──┐\n                              ├──▶  Step 2: Grounding expression  ──▶  [Step 3: Double check]\nStep 1: Image caption  ──────┘                                                   (optional)\n```\n\n- **Step 0 (region_expr)** — VLM emits one short discriminative phrase per KITTI bbox (`bbox_2d`, `type`, `color`, `description`).\n- **Step 1 (image_caption)** — VLM emits a holistic, location-agnostic scene caption.\n- **Step 2 (grounding_expr)** — VLM groups Step 0 objects into grouping phrases and returns one bbox list per group, optionally using Step 1's caption as extra context.\n- **Step 3 (double_check)** — VLM re-checks each Step 2 bbox against the image; bad matches are removed, slightly-off boxes get tightened.\n\nSteps 0 and 1 run in parallel within a single thread pool (they only depend on the seed records). Each step writes its own `step_\u003CN>_*\u002Fannotations.jsonl` and skips already-processed images on re-run unless `workflow.force_reprocess: true`.\n\n## Instructions\n\n### Initial setup\n\nWhen a user wants to run this pipeline, walk through these steps:\n\n1. **Images**: Ask for `data.image_dir`, the directory containing `.jpg`, `.jpeg`, or `.png` images.\n2. **KITTI labels**: Ask for `data.kitti_label_dir`, the directory containing one `.txt` label file per image. Each label line must use KITTI format: `\u003Ctype> \u003Ctruncated> \u003Coccluded> \u003Calpha> \u003Cbbox_left> \u003Cbbox_top> \u003Cbbox_right> \u003Cbbox_bottom> ...`. Lines with fewer than 8 fields are silently skipped. Set this even for Step 1-only runs because Steps 0 and 2 require it.\n3. **Resume from existing annotations**: If the user already has a unified `annotations.jsonl` from a previous run, set `data.input_annotations_jsonl` to that file instead of seeding from `data.image_dir` and `data.kitti_label_dir`.\n4. **API access**: Ask the user which VLM endpoint they want to use. Present these five options and act on the choice:\n   1. **Gemini** — set `vlm.backend: \"gemini\"`; require `GOOGLE_API_KEY` (env var or `vlm.gemini.api_key`).\n   2. **NIM** (e.g. `https:\u002F\u002Finference-api.nvidia.com\u002Fv1`) — set `vlm.backend: \"openai\"`; collect `base_url`, `model_name`, and `api_key`.\n   3. **TAO inference microservice** (self-hosted, OpenAI-compatible). Confirm whether the server is already running:\n      - **Running** — collect `base_url`, `model_name`, and (optionally) `api_key`; set `vlm.backend: \"openai\"`.\n      - **Not running** — guide the user through the `skills\u002Fapplications\u002Ftao-run-inference-service` skill, which stands up a local TAO inference microservice with an OpenAI-compatible API. Before promising a specific model, check `skills\u002Fapplications\u002Ftao-run-inference-service\u002Freferences\u002Fservice.yaml` for `valid_network_arch_config_basenames`. Once the server is up, collect `base_url`, `model_name`, and (optionally) `api_key`; set `vlm.backend: \"openai\"`.\n   4. **vLLM** (self-hosted, OpenAI-compatible). Confirm whether the server is already running:\n      - **Running** — collect `base_url`, `model_name`, and (optionally) `api_key`; set `vlm.backend: \"openai\"`.\n      - **Not running** — follow [references\u002Fvllm_server.md](references\u002Fvllm_server.md) to install and launch a vLLM server, then collect `base_url`, `model_name`, and (optionally) `api_key`; set `vlm.backend: \"openai\"`.\n   5. **Custom** (any other OpenAI-compatible endpoint) — set `vlm.backend: \"openai\"`; collect `base_url`, `model_name`, and (optionally) `api_key`.\n\n   If the user has no endpoint and does not want to set one up, stop and help resolve API access first.\n5. **Workflow steps**: Choose one of:\n   - Full pipeline: `[\"0\", \"1\", \"2\", \"3\"]`\n   - No caption generation: `[\"0\", \"2\", \"3\"]`, where Step 2 falls back to image-only context\n   - No verification: `[\"0\", \"1\", \"2\"]`\n   - Custom subset: any supported subset of steps\n6. **Output format**: Choose one of:\n   - `jsonl`: unified schema only\n   - `legacy`: byte-compatible `.txt.stepN` files only\n   - `both`: writes both formats and is the default for downstream tooling\n\n### Running the pipeline\n\nThe pipeline runs inside the TAO Toolkit container via the `auto_label` CLI:\n\n```bash\nauto_label generate -e \u002Fpath\u002Fto\u002Fspec.yaml \\\n    results_dir=\u002Fresults \\\n    image_referring_expression.data.image_dir=\u002Fdata\u002Fimages \\\n    image_referring_expression.data.kitti_label_dir=\u002Fdata\u002Flabels \\\n    image_referring_expression.vlm.gemini.api_key=$GOOGLE_API_KEY\n```\n\nGenerate a default spec: `auto_label default_specs results_dir=\u002Fresults module_name=auto_label`, then set `autolabel_type: \"image_referring_expression\"`. All fields support Hydra dot-notation overrides on the command line.\n\nSee [references\u002Fconfiguration.md](references\u002Fconfiguration.md) for the full YAML structure, all parameters, model\u002Fendpoint setup, and error patterns.\n\n### Recommended pilot workflow\n\n1. Run on 5-10 images with all four steps.\n2. Inspect `step_0_region_expr\u002Fannotations.jsonl` — are object types, colors, and discriminating phrases accurate?\n3. Inspect `step_2_grounding_expr\u002Fannotations.jsonl` — are objects grouped sensibly, and do bbox coordinates match the described groups?\n4. Inspect `step_3_double_check\u002Fannotations.jsonl` — were mismatched bboxes removed or tightened? Are any new errors introduced (rare)?\n5. If quality is insufficient, switch the VLM to a stronger model (e.g. `gemini-2.5-pro` or a larger Qwen3-VL endpoint), raise `media_resolution` \u002F `max_output_tokens`, then re-run with `workflow.force_reprocess=true`.\n6. Scale to the full dataset once satisfied.\n\n## Configuration\n\nKey configuration fields (full reference in [references\u002Fconfiguration.md](references\u002Fconfiguration.md)):\n\n| Field | Default | Description |\n|-------|---------|-------------|\n| `workflow.steps` | `[\"0\",\"1\",\"2\",\"3\"]` | Which steps to execute (`0`=region_expr, `1`=image_caption, `2`=grounding_expr, `3`=double_check) |\n| `workflow.max_workers` | `4` | Parallel threads per step (watch API rate limits) |\n| `workflow.force_reprocess` | `false` | Ignore cached per-step outputs and reprocess from scratch |\n| `workflow.output_format` | `\"jsonl\"` (set to `\"both\"` in the default spec) | `\"jsonl\"`, `\"legacy\"`, or `\"both\"` |\n| `vlm.backend` | `\"gemini\"` | `\"gemini\"` or `\"openai\"` (OpenAI-compatible endpoint) |\n| `data.image_dir` | required | Directory of input images (`.jpg` \u002F `.jpeg` \u002F `.png`) |\n| `data.kitti_label_dir` | required (unless resuming) | Directory of KITTI-format `.txt` label files |\n| `data.input_annotations_jsonl` | `\"\"` | Optional pre-seeded `annotations.jsonl` (skips KITTI seeding) |\n\n## Inputs\n\nTwo ways to seed the pipeline:\n\n1. **Image directory + KITTI labels** (default). Set `data.image_dir` and `data.kitti_label_dir`. The orchestrator walks the image directory, reads the matching `\u003Cstem>.txt` KITTI file, parses bboxes (fields 0 + 4-7), reads each image's `width`\u002F`height` via PIL, and writes a `seed_annotations.jsonl` to `results_dir\u002F`.\n2. **Pre-seeded annotations JSONL** (resume \u002F pre-computed regions). Set `data.input_annotations_jsonl` to a file with one `{\"image_id\", \"image_path\", \"width\", \"height\", \"kitti_bboxes\": [...]}` object per line.\n\n## Outputs\n\nAll outputs go to `results_dir\u002F`:\n\n- `seed_annotations.jsonl` — initial per-image records (unless `input_annotations_jsonl` was supplied).\n- `step_0_region_expr\u002Fannotations.jsonl` — adds `regions[]` (each with `bbox`\u002F`bbox_2d`, `type`, `color`, `description`).\n- `step_1_image_caption\u002Fannotations.jsonl` — adds `caption` (string).\n- `step_2_grounding_expr\u002Fannotations.jsonl` — adds `expressions[]` (each `{text, instances: [{bbox: [x1,y1,x2,y2]}]}`).\n- `step_3_double_check\u002Fannotations.jsonl` — same shape as Step 2, with bboxes removed\u002Fupdated.\n- `results_dir\u002Fannotations.jsonl` — copy of the last completed step's output.\n- When `workflow.output_format` is `\"legacy\"` or `\"both\"`, each step also writes byte-compatible `step_\u003CN>_*\u002Flabels\u002F\u003Cstem>.txt.stepN` files for the original 2d-data-engine tooling.\n\n## Prerequisites\n\n- **Container**: `nvcr.io\u002Fnvidia\u002Ftao\u002Ftao-toolkit:6.26.3-pyt`\n- **API access**: At least one VLM endpoint (Gemini API key or OpenAI-compatible endpoint capable of image input)\n- **PIL \u002F Pillow**: Required to read image dimensions during seeding (already present in the TAO container)\n",{"data":32,"body":45},{"name":4,"description":6,"license":23,"compatibility":33,"metadata":34,"tags":37,"allowed-tools":44},"Requires docker + nvidia-container-toolkit + at least one VLM endpoint (Gemini API key or OpenAI-compatible).",{"author":35,"version":36},"NVIDIA Corporation","0.1.0",[38,39,40,41,42,43],"image","referring-expression","kitti","bounding-boxes","auto-label","vlm","Read Bash Write",{"type":46,"children":47},"root",[48,57,63,70,92,98,110,186,207,213,220,225,773,779,792,887,908,920,926,1012,1018,1029,1363,1369,1374,1464,1470,1482,1648,1654,1692],{"type":49,"tag":50,"props":51,"children":53},"element","h1",{"id":52},"image-referring-expression-pipeline",[54],{"type":55,"value":56},"text","Image Referring Expression Pipeline",{"type":49,"tag":58,"props":59,"children":60},"p",{},[61],{"type":55,"value":62},"Generate referring-expression and grounding annotations from images with KITTI-format bounding box labels. A single VLM (Gemini or any OpenAI-compatible endpoint) runs four steps: per-object region descriptions, holistic image captions, grouped grounding expressions tied to bboxes, and an optional double-check verification pass.",{"type":49,"tag":64,"props":65,"children":67},"h2",{"id":66},"purpose",[68],{"type":55,"value":69},"Purpose",{"type":49,"tag":58,"props":71,"children":72},{},[73,75,82,84,90],{"type":55,"value":74},"Transform ",{"type":49,"tag":76,"props":77,"children":79},"code",{"className":78},[],[80],{"type":55,"value":81},"(image, KITTI labels)",{"type":55,"value":83}," pairs into a unified ",{"type":49,"tag":76,"props":85,"children":87},{"className":86},[],[88],{"type":55,"value":89},"annotations.jsonl",{"type":55,"value":91}," containing rich, grounded referring expressions. The VLM acts as a \"teacher\" annotator: Steps 0-1 see the image; Step 2 groups Step 0 outputs into grouping phrases with bbox lists; Step 3 (optional) re-examines those bboxes against the image and corrects mismatches.",{"type":49,"tag":64,"props":93,"children":95},{"id":94},"pipeline-architecture",[96],{"type":55,"value":97},"Pipeline Architecture",{"type":49,"tag":99,"props":100,"children":104},"pre",{"className":101,"code":103,"language":55},[102],"language-text","Step 0: Region expression  ──┐\n                              ├──▶  Step 2: Grounding expression  ──▶  [Step 3: Double check]\nStep 1: Image caption  ──────┘                                                   (optional)\n",[105],{"type":49,"tag":76,"props":106,"children":108},{"__ignoreMap":107},"",[109],{"type":55,"value":103},{"type":49,"tag":111,"props":112,"children":113},"ul",{},[114,156,166,176],{"type":49,"tag":115,"props":116,"children":117},"li",{},[118,124,126,132,134,140,141,147,148,154],{"type":49,"tag":119,"props":120,"children":121},"strong",{},[122],{"type":55,"value":123},"Step 0 (region_expr)",{"type":55,"value":125}," — VLM emits one short discriminative phrase per KITTI bbox (",{"type":49,"tag":76,"props":127,"children":129},{"className":128},[],[130],{"type":55,"value":131},"bbox_2d",{"type":55,"value":133},", ",{"type":49,"tag":76,"props":135,"children":137},{"className":136},[],[138],{"type":55,"value":139},"type",{"type":55,"value":133},{"type":49,"tag":76,"props":142,"children":144},{"className":143},[],[145],{"type":55,"value":146},"color",{"type":55,"value":133},{"type":49,"tag":76,"props":149,"children":151},{"className":150},[],[152],{"type":55,"value":153},"description",{"type":55,"value":155},").",{"type":49,"tag":115,"props":157,"children":158},{},[159,164],{"type":49,"tag":119,"props":160,"children":161},{},[162],{"type":55,"value":163},"Step 1 (image_caption)",{"type":55,"value":165}," — VLM emits a holistic, location-agnostic scene caption.",{"type":49,"tag":115,"props":167,"children":168},{},[169,174],{"type":49,"tag":119,"props":170,"children":171},{},[172],{"type":55,"value":173},"Step 2 (grounding_expr)",{"type":55,"value":175}," — VLM groups Step 0 objects into grouping phrases and returns one bbox list per group, optionally using Step 1's caption as extra context.",{"type":49,"tag":115,"props":177,"children":178},{},[179,184],{"type":49,"tag":119,"props":180,"children":181},{},[182],{"type":55,"value":183},"Step 3 (double_check)",{"type":55,"value":185}," — VLM re-checks each Step 2 bbox against the image; bad matches are removed, slightly-off boxes get tightened.",{"type":49,"tag":58,"props":187,"children":188},{},[189,191,197,199,205],{"type":55,"value":190},"Steps 0 and 1 run in parallel within a single thread pool (they only depend on the seed records). Each step writes its own ",{"type":49,"tag":76,"props":192,"children":194},{"className":193},[],[195],{"type":55,"value":196},"step_\u003CN>_*\u002Fannotations.jsonl",{"type":55,"value":198}," and skips already-processed images on re-run unless ",{"type":49,"tag":76,"props":200,"children":202},{"className":201},[],[203],{"type":55,"value":204},"workflow.force_reprocess: true",{"type":55,"value":206},".",{"type":49,"tag":64,"props":208,"children":210},{"id":209},"instructions",[211],{"type":55,"value":212},"Instructions",{"type":49,"tag":214,"props":215,"children":217},"h3",{"id":216},"initial-setup",[218],{"type":55,"value":219},"Initial setup",{"type":49,"tag":58,"props":221,"children":222},{},[223],{"type":55,"value":224},"When a user wants to run this pipeline, walk through these steps:",{"type":49,"tag":226,"props":227,"children":228},"ol",{},[229,270,303,341,667,720],{"type":49,"tag":115,"props":230,"children":231},{},[232,237,239,245,247,253,254,260,262,268],{"type":49,"tag":119,"props":233,"children":234},{},[235],{"type":55,"value":236},"Images",{"type":55,"value":238},": Ask for ",{"type":49,"tag":76,"props":240,"children":242},{"className":241},[],[243],{"type":55,"value":244},"data.image_dir",{"type":55,"value":246},", the directory containing ",{"type":49,"tag":76,"props":248,"children":250},{"className":249},[],[251],{"type":55,"value":252},".jpg",{"type":55,"value":133},{"type":49,"tag":76,"props":255,"children":257},{"className":256},[],[258],{"type":55,"value":259},".jpeg",{"type":55,"value":261},", or ",{"type":49,"tag":76,"props":263,"children":265},{"className":264},[],[266],{"type":55,"value":267},".png",{"type":55,"value":269}," images.",{"type":49,"tag":115,"props":271,"children":272},{},[273,278,279,285,287,293,295,301],{"type":49,"tag":119,"props":274,"children":275},{},[276],{"type":55,"value":277},"KITTI labels",{"type":55,"value":238},{"type":49,"tag":76,"props":280,"children":282},{"className":281},[],[283],{"type":55,"value":284},"data.kitti_label_dir",{"type":55,"value":286},", the directory containing one ",{"type":49,"tag":76,"props":288,"children":290},{"className":289},[],[291],{"type":55,"value":292},".txt",{"type":55,"value":294}," label file per image. Each label line must use KITTI format: ",{"type":49,"tag":76,"props":296,"children":298},{"className":297},[],[299],{"type":55,"value":300},"\u003Ctype> \u003Ctruncated> \u003Coccluded> \u003Calpha> \u003Cbbox_left> \u003Cbbox_top> \u003Cbbox_right> \u003Cbbox_bottom> ...",{"type":55,"value":302},". Lines with fewer than 8 fields are silently skipped. Set this even for Step 1-only runs because Steps 0 and 2 require it.",{"type":49,"tag":115,"props":304,"children":305},{},[306,311,313,318,320,326,328,333,335,340],{"type":49,"tag":119,"props":307,"children":308},{},[309],{"type":55,"value":310},"Resume from existing annotations",{"type":55,"value":312},": If the user already has a unified ",{"type":49,"tag":76,"props":314,"children":316},{"className":315},[],[317],{"type":55,"value":89},{"type":55,"value":319}," from a previous run, set ",{"type":49,"tag":76,"props":321,"children":323},{"className":322},[],[324],{"type":55,"value":325},"data.input_annotations_jsonl",{"type":55,"value":327}," to that file instead of seeding from ",{"type":49,"tag":76,"props":329,"children":331},{"className":330},[],[332],{"type":55,"value":244},{"type":55,"value":334}," and ",{"type":49,"tag":76,"props":336,"children":338},{"className":337},[],[339],{"type":55,"value":284},{"type":55,"value":206},{"type":49,"tag":115,"props":342,"children":343},{},[344,349,351,661,665],{"type":49,"tag":119,"props":345,"children":346},{},[347],{"type":55,"value":348},"API access",{"type":55,"value":350},": Ask the user which VLM endpoint they want to use. Present these five options and act on the choice:",{"type":49,"tag":226,"props":352,"children":353},{},[354,387,435,542,627],{"type":49,"tag":115,"props":355,"children":356},{},[357,362,364,370,372,378,380,386],{"type":49,"tag":119,"props":358,"children":359},{},[360],{"type":55,"value":361},"Gemini",{"type":55,"value":363}," — set ",{"type":49,"tag":76,"props":365,"children":367},{"className":366},[],[368],{"type":55,"value":369},"vlm.backend: \"gemini\"",{"type":55,"value":371},"; require ",{"type":49,"tag":76,"props":373,"children":375},{"className":374},[],[376],{"type":55,"value":377},"GOOGLE_API_KEY",{"type":55,"value":379}," (env var or ",{"type":49,"tag":76,"props":381,"children":383},{"className":382},[],[384],{"type":55,"value":385},"vlm.gemini.api_key",{"type":55,"value":155},{"type":49,"tag":115,"props":388,"children":389},{},[390,395,397,403,405,411,413,419,420,426,428,434],{"type":49,"tag":119,"props":391,"children":392},{},[393],{"type":55,"value":394},"NIM",{"type":55,"value":396}," (e.g. ",{"type":49,"tag":76,"props":398,"children":400},{"className":399},[],[401],{"type":55,"value":402},"https:\u002F\u002Finference-api.nvidia.com\u002Fv1",{"type":55,"value":404},") — set ",{"type":49,"tag":76,"props":406,"children":408},{"className":407},[],[409],{"type":55,"value":410},"vlm.backend: \"openai\"",{"type":55,"value":412},"; collect ",{"type":49,"tag":76,"props":414,"children":416},{"className":415},[],[417],{"type":55,"value":418},"base_url",{"type":55,"value":133},{"type":49,"tag":76,"props":421,"children":423},{"className":422},[],[424],{"type":55,"value":425},"model_name",{"type":55,"value":427},", and ",{"type":49,"tag":76,"props":429,"children":431},{"className":430},[],[432],{"type":55,"value":433},"api_key",{"type":55,"value":206},{"type":49,"tag":115,"props":436,"children":437},{},[438,443,445],{"type":49,"tag":119,"props":439,"children":440},{},[441],{"type":55,"value":442},"TAO inference microservice",{"type":55,"value":444}," (self-hosted, OpenAI-compatible). Confirm whether the server is already running:\n",{"type":49,"tag":111,"props":446,"children":447},{},[448,484],{"type":49,"tag":115,"props":449,"children":450},{},[451,456,458,463,464,469,471,476,478,483],{"type":49,"tag":119,"props":452,"children":453},{},[454],{"type":55,"value":455},"Running",{"type":55,"value":457}," — collect ",{"type":49,"tag":76,"props":459,"children":461},{"className":460},[],[462],{"type":55,"value":418},{"type":55,"value":133},{"type":49,"tag":76,"props":465,"children":467},{"className":466},[],[468],{"type":55,"value":425},{"type":55,"value":470},", and (optionally) ",{"type":49,"tag":76,"props":472,"children":474},{"className":473},[],[475],{"type":55,"value":433},{"type":55,"value":477},"; set ",{"type":49,"tag":76,"props":479,"children":481},{"className":480},[],[482],{"type":55,"value":410},{"type":55,"value":206},{"type":49,"tag":115,"props":485,"children":486},{},[487,492,494,500,502,508,510,516,518,523,524,529,530,535,536,541],{"type":49,"tag":119,"props":488,"children":489},{},[490],{"type":55,"value":491},"Not running",{"type":55,"value":493}," — guide the user through the ",{"type":49,"tag":76,"props":495,"children":497},{"className":496},[],[498],{"type":55,"value":499},"skills\u002Fapplications\u002Ftao-run-inference-service",{"type":55,"value":501}," skill, which stands up a local TAO inference microservice with an OpenAI-compatible API. Before promising a specific model, check ",{"type":49,"tag":76,"props":503,"children":505},{"className":504},[],[506],{"type":55,"value":507},"skills\u002Fapplications\u002Ftao-run-inference-service\u002Freferences\u002Fservice.yaml",{"type":55,"value":509}," for ",{"type":49,"tag":76,"props":511,"children":513},{"className":512},[],[514],{"type":55,"value":515},"valid_network_arch_config_basenames",{"type":55,"value":517},". Once the server is up, collect ",{"type":49,"tag":76,"props":519,"children":521},{"className":520},[],[522],{"type":55,"value":418},{"type":55,"value":133},{"type":49,"tag":76,"props":525,"children":527},{"className":526},[],[528],{"type":55,"value":425},{"type":55,"value":470},{"type":49,"tag":76,"props":531,"children":533},{"className":532},[],[534],{"type":55,"value":433},{"type":55,"value":477},{"type":49,"tag":76,"props":537,"children":539},{"className":538},[],[540],{"type":55,"value":410},{"type":55,"value":206},{"type":49,"tag":115,"props":543,"children":544},{},[545,550,551],{"type":49,"tag":119,"props":546,"children":547},{},[548],{"type":55,"value":549},"vLLM",{"type":55,"value":444},{"type":49,"tag":111,"props":552,"children":553},{},[554,586],{"type":49,"tag":115,"props":555,"children":556},{},[557,561,562,567,568,573,574,579,580,585],{"type":49,"tag":119,"props":558,"children":559},{},[560],{"type":55,"value":455},{"type":55,"value":457},{"type":49,"tag":76,"props":563,"children":565},{"className":564},[],[566],{"type":55,"value":418},{"type":55,"value":133},{"type":49,"tag":76,"props":569,"children":571},{"className":570},[],[572],{"type":55,"value":425},{"type":55,"value":470},{"type":49,"tag":76,"props":575,"children":577},{"className":576},[],[578],{"type":55,"value":433},{"type":55,"value":477},{"type":49,"tag":76,"props":581,"children":583},{"className":582},[],[584],{"type":55,"value":410},{"type":55,"value":206},{"type":49,"tag":115,"props":587,"children":588},{},[589,593,595,601,603,608,609,614,615,620,621,626],{"type":49,"tag":119,"props":590,"children":591},{},[592],{"type":55,"value":491},{"type":55,"value":594}," — follow ",{"type":49,"tag":596,"props":597,"children":599},"a",{"href":598},"references\u002Fvllm_server.md",[600],{"type":55,"value":598},{"type":55,"value":602}," to install and launch a vLLM server, then collect ",{"type":49,"tag":76,"props":604,"children":606},{"className":605},[],[607],{"type":55,"value":418},{"type":55,"value":133},{"type":49,"tag":76,"props":610,"children":612},{"className":611},[],[613],{"type":55,"value":425},{"type":55,"value":470},{"type":49,"tag":76,"props":616,"children":618},{"className":617},[],[619],{"type":55,"value":433},{"type":55,"value":477},{"type":49,"tag":76,"props":622,"children":624},{"className":623},[],[625],{"type":55,"value":410},{"type":55,"value":206},{"type":49,"tag":115,"props":628,"children":629},{},[630,635,637,642,643,648,649,654,655,660],{"type":49,"tag":119,"props":631,"children":632},{},[633],{"type":55,"value":634},"Custom",{"type":55,"value":636}," (any other OpenAI-compatible endpoint) — set ",{"type":49,"tag":76,"props":638,"children":640},{"className":639},[],[641],{"type":55,"value":410},{"type":55,"value":412},{"type":49,"tag":76,"props":644,"children":646},{"className":645},[],[647],{"type":55,"value":418},{"type":55,"value":133},{"type":49,"tag":76,"props":650,"children":652},{"className":651},[],[653],{"type":55,"value":425},{"type":55,"value":470},{"type":49,"tag":76,"props":656,"children":658},{"className":657},[],[659],{"type":55,"value":433},{"type":55,"value":206},{"type":49,"tag":662,"props":663,"children":664},"br",{},[],{"type":55,"value":666},"If the user has no endpoint and does not want to set one up, stop and help resolve API access first.",{"type":49,"tag":115,"props":668,"children":669},{},[670,675,677],{"type":49,"tag":119,"props":671,"children":672},{},[673],{"type":55,"value":674},"Workflow steps",{"type":55,"value":676},": Choose one of:",{"type":49,"tag":111,"props":678,"children":679},{},[680,691,704,715],{"type":49,"tag":115,"props":681,"children":682},{},[683,685],{"type":55,"value":684},"Full pipeline: ",{"type":49,"tag":76,"props":686,"children":688},{"className":687},[],[689],{"type":55,"value":690},"[\"0\", \"1\", \"2\", \"3\"]",{"type":49,"tag":115,"props":692,"children":693},{},[694,696,702],{"type":55,"value":695},"No caption generation: ",{"type":49,"tag":76,"props":697,"children":699},{"className":698},[],[700],{"type":55,"value":701},"[\"0\", \"2\", \"3\"]",{"type":55,"value":703},", where Step 2 falls back to image-only context",{"type":49,"tag":115,"props":705,"children":706},{},[707,709],{"type":55,"value":708},"No verification: ",{"type":49,"tag":76,"props":710,"children":712},{"className":711},[],[713],{"type":55,"value":714},"[\"0\", \"1\", \"2\"]",{"type":49,"tag":115,"props":716,"children":717},{},[718],{"type":55,"value":719},"Custom subset: any supported subset of steps",{"type":49,"tag":115,"props":721,"children":722},{},[723,728,729],{"type":49,"tag":119,"props":724,"children":725},{},[726],{"type":55,"value":727},"Output format",{"type":55,"value":676},{"type":49,"tag":111,"props":730,"children":731},{},[732,743,762],{"type":49,"tag":115,"props":733,"children":734},{},[735,741],{"type":49,"tag":76,"props":736,"children":738},{"className":737},[],[739],{"type":55,"value":740},"jsonl",{"type":55,"value":742},": unified schema only",{"type":49,"tag":115,"props":744,"children":745},{},[746,752,754,760],{"type":49,"tag":76,"props":747,"children":749},{"className":748},[],[750],{"type":55,"value":751},"legacy",{"type":55,"value":753},": byte-compatible ",{"type":49,"tag":76,"props":755,"children":757},{"className":756},[],[758],{"type":55,"value":759},".txt.stepN",{"type":55,"value":761}," files only",{"type":49,"tag":115,"props":763,"children":764},{},[765,771],{"type":49,"tag":76,"props":766,"children":768},{"className":767},[],[769],{"type":55,"value":770},"both",{"type":55,"value":772},": writes both formats and is the default for downstream tooling",{"type":49,"tag":214,"props":774,"children":776},{"id":775},"running-the-pipeline",[777],{"type":55,"value":778},"Running the pipeline",{"type":49,"tag":58,"props":780,"children":781},{},[782,784,790],{"type":55,"value":783},"The pipeline runs inside the TAO Toolkit container via the ",{"type":49,"tag":76,"props":785,"children":787},{"className":786},[],[788],{"type":55,"value":789},"auto_label",{"type":55,"value":791}," CLI:",{"type":49,"tag":99,"props":793,"children":797},{"className":794,"code":795,"language":796,"meta":107,"style":107},"language-bash shiki shiki-themes material-theme-lighter material-theme material-theme-palenight","auto_label generate -e \u002Fpath\u002Fto\u002Fspec.yaml \\\n    results_dir=\u002Fresults \\\n    image_referring_expression.data.image_dir=\u002Fdata\u002Fimages \\\n    image_referring_expression.data.kitti_label_dir=\u002Fdata\u002Flabels \\\n    image_referring_expression.vlm.gemini.api_key=$GOOGLE_API_KEY\n","bash",[798],{"type":49,"tag":76,"props":799,"children":800},{"__ignoreMap":107},[801,834,847,860,873],{"type":49,"tag":802,"props":803,"children":806},"span",{"class":804,"line":805},"line",1,[807,812,818,823,828],{"type":49,"tag":802,"props":808,"children":810},{"style":809},"--shiki-light:#E2931D;--shiki-default:#FFCB6B;--shiki-dark:#FFCB6B",[811],{"type":55,"value":789},{"type":49,"tag":802,"props":813,"children":815},{"style":814},"--shiki-light:#91B859;--shiki-default:#C3E88D;--shiki-dark:#C3E88D",[816],{"type":55,"value":817}," generate",{"type":49,"tag":802,"props":819,"children":820},{"style":814},[821],{"type":55,"value":822}," -e",{"type":49,"tag":802,"props":824,"children":825},{"style":814},[826],{"type":55,"value":827}," \u002Fpath\u002Fto\u002Fspec.yaml",{"type":49,"tag":802,"props":829,"children":831},{"style":830},"--shiki-light:#90A4AE;--shiki-default:#EEFFFF;--shiki-dark:#BABED8",[832],{"type":55,"value":833}," \\\n",{"type":49,"tag":802,"props":835,"children":837},{"class":804,"line":836},2,[838,843],{"type":49,"tag":802,"props":839,"children":840},{"style":814},[841],{"type":55,"value":842},"    results_dir=\u002Fresults",{"type":49,"tag":802,"props":844,"children":845},{"style":830},[846],{"type":55,"value":833},{"type":49,"tag":802,"props":848,"children":850},{"class":804,"line":849},3,[851,856],{"type":49,"tag":802,"props":852,"children":853},{"style":814},[854],{"type":55,"value":855},"    image_referring_expression.data.image_dir=\u002Fdata\u002Fimages",{"type":49,"tag":802,"props":857,"children":858},{"style":830},[859],{"type":55,"value":833},{"type":49,"tag":802,"props":861,"children":863},{"class":804,"line":862},4,[864,869],{"type":49,"tag":802,"props":865,"children":866},{"style":814},[867],{"type":55,"value":868},"    image_referring_expression.data.kitti_label_dir=\u002Fdata\u002Flabels",{"type":49,"tag":802,"props":870,"children":871},{"style":830},[872],{"type":55,"value":833},{"type":49,"tag":802,"props":874,"children":876},{"class":804,"line":875},5,[877,882],{"type":49,"tag":802,"props":878,"children":879},{"style":814},[880],{"type":55,"value":881},"    image_referring_expression.vlm.gemini.api_key=",{"type":49,"tag":802,"props":883,"children":884},{"style":830},[885],{"type":55,"value":886},"$GOOGLE_API_KEY\n",{"type":49,"tag":58,"props":888,"children":889},{},[890,892,898,900,906],{"type":55,"value":891},"Generate a default spec: ",{"type":49,"tag":76,"props":893,"children":895},{"className":894},[],[896],{"type":55,"value":897},"auto_label default_specs results_dir=\u002Fresults module_name=auto_label",{"type":55,"value":899},", then set ",{"type":49,"tag":76,"props":901,"children":903},{"className":902},[],[904],{"type":55,"value":905},"autolabel_type: \"image_referring_expression\"",{"type":55,"value":907},". All fields support Hydra dot-notation overrides on the command line.",{"type":49,"tag":58,"props":909,"children":910},{},[911,913,918],{"type":55,"value":912},"See ",{"type":49,"tag":596,"props":914,"children":916},{"href":915},"references\u002Fconfiguration.md",[917],{"type":55,"value":915},{"type":55,"value":919}," for the full YAML structure, all parameters, model\u002Fendpoint setup, and error patterns.",{"type":49,"tag":214,"props":921,"children":923},{"id":922},"recommended-pilot-workflow",[924],{"type":55,"value":925},"Recommended pilot workflow",{"type":49,"tag":226,"props":927,"children":928},{},[929,934,947,959,971,1007],{"type":49,"tag":115,"props":930,"children":931},{},[932],{"type":55,"value":933},"Run on 5-10 images with all four steps.",{"type":49,"tag":115,"props":935,"children":936},{},[937,939,945],{"type":55,"value":938},"Inspect ",{"type":49,"tag":76,"props":940,"children":942},{"className":941},[],[943],{"type":55,"value":944},"step_0_region_expr\u002Fannotations.jsonl",{"type":55,"value":946}," — are object types, colors, and discriminating phrases accurate?",{"type":49,"tag":115,"props":948,"children":949},{},[950,951,957],{"type":55,"value":938},{"type":49,"tag":76,"props":952,"children":954},{"className":953},[],[955],{"type":55,"value":956},"step_2_grounding_expr\u002Fannotations.jsonl",{"type":55,"value":958}," — are objects grouped sensibly, and do bbox coordinates match the described groups?",{"type":49,"tag":115,"props":960,"children":961},{},[962,963,969],{"type":55,"value":938},{"type":49,"tag":76,"props":964,"children":966},{"className":965},[],[967],{"type":55,"value":968},"step_3_double_check\u002Fannotations.jsonl",{"type":55,"value":970}," — were mismatched bboxes removed or tightened? Are any new errors introduced (rare)?",{"type":49,"tag":115,"props":972,"children":973},{},[974,976,982,984,990,992,998,1000,1006],{"type":55,"value":975},"If quality is insufficient, switch the VLM to a stronger model (e.g. ",{"type":49,"tag":76,"props":977,"children":979},{"className":978},[],[980],{"type":55,"value":981},"gemini-2.5-pro",{"type":55,"value":983}," or a larger Qwen3-VL endpoint), raise ",{"type":49,"tag":76,"props":985,"children":987},{"className":986},[],[988],{"type":55,"value":989},"media_resolution",{"type":55,"value":991}," \u002F ",{"type":49,"tag":76,"props":993,"children":995},{"className":994},[],[996],{"type":55,"value":997},"max_output_tokens",{"type":55,"value":999},", then re-run with ",{"type":49,"tag":76,"props":1001,"children":1003},{"className":1002},[],[1004],{"type":55,"value":1005},"workflow.force_reprocess=true",{"type":55,"value":206},{"type":49,"tag":115,"props":1008,"children":1009},{},[1010],{"type":55,"value":1011},"Scale to the full dataset once satisfied.",{"type":49,"tag":64,"props":1013,"children":1015},{"id":1014},"configuration",[1016],{"type":55,"value":1017},"Configuration",{"type":49,"tag":58,"props":1019,"children":1020},{},[1021,1023,1027],{"type":55,"value":1022},"Key configuration fields (full reference in ",{"type":49,"tag":596,"props":1024,"children":1025},{"href":915},[1026],{"type":55,"value":915},{"type":55,"value":1028},"):",{"type":49,"tag":1030,"props":1031,"children":1032},"table",{},[1033,1057],{"type":49,"tag":1034,"props":1035,"children":1036},"thead",{},[1037],{"type":49,"tag":1038,"props":1039,"children":1040},"tr",{},[1041,1047,1052],{"type":49,"tag":1042,"props":1043,"children":1044},"th",{},[1045],{"type":55,"value":1046},"Field",{"type":49,"tag":1042,"props":1048,"children":1049},{},[1050],{"type":55,"value":1051},"Default",{"type":49,"tag":1042,"props":1053,"children":1054},{},[1055],{"type":55,"value":1056},"Description",{"type":49,"tag":1058,"props":1059,"children":1060},"tbody",{},[1061,1120,1146,1172,1224,1263,1303,1331],{"type":49,"tag":1038,"props":1062,"children":1063},{},[1064,1074,1083],{"type":49,"tag":1065,"props":1066,"children":1067},"td",{},[1068],{"type":49,"tag":76,"props":1069,"children":1071},{"className":1070},[],[1072],{"type":55,"value":1073},"workflow.steps",{"type":49,"tag":1065,"props":1075,"children":1076},{},[1077],{"type":49,"tag":76,"props":1078,"children":1080},{"className":1079},[],[1081],{"type":55,"value":1082},"[\"0\",\"1\",\"2\",\"3\"]",{"type":49,"tag":1065,"props":1084,"children":1085},{},[1086,1088,1094,1096,1102,1104,1110,1112,1118],{"type":55,"value":1087},"Which steps to execute (",{"type":49,"tag":76,"props":1089,"children":1091},{"className":1090},[],[1092],{"type":55,"value":1093},"0",{"type":55,"value":1095},"=region_expr, ",{"type":49,"tag":76,"props":1097,"children":1099},{"className":1098},[],[1100],{"type":55,"value":1101},"1",{"type":55,"value":1103},"=image_caption, ",{"type":49,"tag":76,"props":1105,"children":1107},{"className":1106},[],[1108],{"type":55,"value":1109},"2",{"type":55,"value":1111},"=grounding_expr, ",{"type":49,"tag":76,"props":1113,"children":1115},{"className":1114},[],[1116],{"type":55,"value":1117},"3",{"type":55,"value":1119},"=double_check)",{"type":49,"tag":1038,"props":1121,"children":1122},{},[1123,1132,1141],{"type":49,"tag":1065,"props":1124,"children":1125},{},[1126],{"type":49,"tag":76,"props":1127,"children":1129},{"className":1128},[],[1130],{"type":55,"value":1131},"workflow.max_workers",{"type":49,"tag":1065,"props":1133,"children":1134},{},[1135],{"type":49,"tag":76,"props":1136,"children":1138},{"className":1137},[],[1139],{"type":55,"value":1140},"4",{"type":49,"tag":1065,"props":1142,"children":1143},{},[1144],{"type":55,"value":1145},"Parallel threads per step (watch API rate limits)",{"type":49,"tag":1038,"props":1147,"children":1148},{},[1149,1158,1167],{"type":49,"tag":1065,"props":1150,"children":1151},{},[1152],{"type":49,"tag":76,"props":1153,"children":1155},{"className":1154},[],[1156],{"type":55,"value":1157},"workflow.force_reprocess",{"type":49,"tag":1065,"props":1159,"children":1160},{},[1161],{"type":49,"tag":76,"props":1162,"children":1164},{"className":1163},[],[1165],{"type":55,"value":1166},"false",{"type":49,"tag":1065,"props":1168,"children":1169},{},[1170],{"type":55,"value":1171},"Ignore cached per-step outputs and reprocess from scratch",{"type":49,"tag":1038,"props":1173,"children":1174},{},[1175,1184,1203],{"type":49,"tag":1065,"props":1176,"children":1177},{},[1178],{"type":49,"tag":76,"props":1179,"children":1181},{"className":1180},[],[1182],{"type":55,"value":1183},"workflow.output_format",{"type":49,"tag":1065,"props":1185,"children":1186},{},[1187,1193,1195,1201],{"type":49,"tag":76,"props":1188,"children":1190},{"className":1189},[],[1191],{"type":55,"value":1192},"\"jsonl\"",{"type":55,"value":1194}," (set to ",{"type":49,"tag":76,"props":1196,"children":1198},{"className":1197},[],[1199],{"type":55,"value":1200},"\"both\"",{"type":55,"value":1202}," in the default spec)",{"type":49,"tag":1065,"props":1204,"children":1205},{},[1206,1211,1212,1218,1219],{"type":49,"tag":76,"props":1207,"children":1209},{"className":1208},[],[1210],{"type":55,"value":1192},{"type":55,"value":133},{"type":49,"tag":76,"props":1213,"children":1215},{"className":1214},[],[1216],{"type":55,"value":1217},"\"legacy\"",{"type":55,"value":261},{"type":49,"tag":76,"props":1220,"children":1222},{"className":1221},[],[1223],{"type":55,"value":1200},{"type":49,"tag":1038,"props":1225,"children":1226},{},[1227,1236,1245],{"type":49,"tag":1065,"props":1228,"children":1229},{},[1230],{"type":49,"tag":76,"props":1231,"children":1233},{"className":1232},[],[1234],{"type":55,"value":1235},"vlm.backend",{"type":49,"tag":1065,"props":1237,"children":1238},{},[1239],{"type":49,"tag":76,"props":1240,"children":1242},{"className":1241},[],[1243],{"type":55,"value":1244},"\"gemini\"",{"type":49,"tag":1065,"props":1246,"children":1247},{},[1248,1253,1255,1261],{"type":49,"tag":76,"props":1249,"children":1251},{"className":1250},[],[1252],{"type":55,"value":1244},{"type":55,"value":1254}," or ",{"type":49,"tag":76,"props":1256,"children":1258},{"className":1257},[],[1259],{"type":55,"value":1260},"\"openai\"",{"type":55,"value":1262}," (OpenAI-compatible endpoint)",{"type":49,"tag":1038,"props":1264,"children":1265},{},[1266,1274,1279],{"type":49,"tag":1065,"props":1267,"children":1268},{},[1269],{"type":49,"tag":76,"props":1270,"children":1272},{"className":1271},[],[1273],{"type":55,"value":244},{"type":49,"tag":1065,"props":1275,"children":1276},{},[1277],{"type":55,"value":1278},"required",{"type":49,"tag":1065,"props":1280,"children":1281},{},[1282,1284,1289,1290,1295,1296,1301],{"type":55,"value":1283},"Directory of input images (",{"type":49,"tag":76,"props":1285,"children":1287},{"className":1286},[],[1288],{"type":55,"value":252},{"type":55,"value":991},{"type":49,"tag":76,"props":1291,"children":1293},{"className":1292},[],[1294],{"type":55,"value":259},{"type":55,"value":991},{"type":49,"tag":76,"props":1297,"children":1299},{"className":1298},[],[1300],{"type":55,"value":267},{"type":55,"value":1302},")",{"type":49,"tag":1038,"props":1304,"children":1305},{},[1306,1314,1319],{"type":49,"tag":1065,"props":1307,"children":1308},{},[1309],{"type":49,"tag":76,"props":1310,"children":1312},{"className":1311},[],[1313],{"type":55,"value":284},{"type":49,"tag":1065,"props":1315,"children":1316},{},[1317],{"type":55,"value":1318},"required (unless resuming)",{"type":49,"tag":1065,"props":1320,"children":1321},{},[1322,1324,1329],{"type":55,"value":1323},"Directory of KITTI-format ",{"type":49,"tag":76,"props":1325,"children":1327},{"className":1326},[],[1328],{"type":55,"value":292},{"type":55,"value":1330}," label files",{"type":49,"tag":1038,"props":1332,"children":1333},{},[1334,1342,1351],{"type":49,"tag":1065,"props":1335,"children":1336},{},[1337],{"type":49,"tag":76,"props":1338,"children":1340},{"className":1339},[],[1341],{"type":55,"value":325},{"type":49,"tag":1065,"props":1343,"children":1344},{},[1345],{"type":49,"tag":76,"props":1346,"children":1348},{"className":1347},[],[1349],{"type":55,"value":1350},"\"\"",{"type":49,"tag":1065,"props":1352,"children":1353},{},[1354,1356,1361],{"type":55,"value":1355},"Optional pre-seeded ",{"type":49,"tag":76,"props":1357,"children":1359},{"className":1358},[],[1360],{"type":55,"value":89},{"type":55,"value":1362}," (skips KITTI seeding)",{"type":49,"tag":64,"props":1364,"children":1366},{"id":1365},"inputs",[1367],{"type":55,"value":1368},"Inputs",{"type":49,"tag":58,"props":1370,"children":1371},{},[1372],{"type":55,"value":1373},"Two ways to seed the pipeline:",{"type":49,"tag":226,"props":1375,"children":1376},{},[1377,1439],{"type":49,"tag":115,"props":1378,"children":1379},{},[1380,1385,1387,1392,1393,1398,1400,1406,1408,1414,1416,1422,1424,1430,1432,1438],{"type":49,"tag":119,"props":1381,"children":1382},{},[1383],{"type":55,"value":1384},"Image directory + KITTI labels",{"type":55,"value":1386}," (default). Set ",{"type":49,"tag":76,"props":1388,"children":1390},{"className":1389},[],[1391],{"type":55,"value":244},{"type":55,"value":334},{"type":49,"tag":76,"props":1394,"children":1396},{"className":1395},[],[1397],{"type":55,"value":284},{"type":55,"value":1399},". The orchestrator walks the image directory, reads the matching ",{"type":49,"tag":76,"props":1401,"children":1403},{"className":1402},[],[1404],{"type":55,"value":1405},"\u003Cstem>.txt",{"type":55,"value":1407}," KITTI file, parses bboxes (fields 0 + 4-7), reads each image's ",{"type":49,"tag":76,"props":1409,"children":1411},{"className":1410},[],[1412],{"type":55,"value":1413},"width",{"type":55,"value":1415},"\u002F",{"type":49,"tag":76,"props":1417,"children":1419},{"className":1418},[],[1420],{"type":55,"value":1421},"height",{"type":55,"value":1423}," via PIL, and writes a ",{"type":49,"tag":76,"props":1425,"children":1427},{"className":1426},[],[1428],{"type":55,"value":1429},"seed_annotations.jsonl",{"type":55,"value":1431}," to ",{"type":49,"tag":76,"props":1433,"children":1435},{"className":1434},[],[1436],{"type":55,"value":1437},"results_dir\u002F",{"type":55,"value":206},{"type":49,"tag":115,"props":1440,"children":1441},{},[1442,1447,1449,1454,1456,1462],{"type":49,"tag":119,"props":1443,"children":1444},{},[1445],{"type":55,"value":1446},"Pre-seeded annotations JSONL",{"type":55,"value":1448}," (resume \u002F pre-computed regions). Set ",{"type":49,"tag":76,"props":1450,"children":1452},{"className":1451},[],[1453],{"type":55,"value":325},{"type":55,"value":1455}," to a file with one ",{"type":49,"tag":76,"props":1457,"children":1459},{"className":1458},[],[1460],{"type":55,"value":1461},"{\"image_id\", \"image_path\", \"width\", \"height\", \"kitti_bboxes\": [...]}",{"type":55,"value":1463}," object per line.",{"type":49,"tag":64,"props":1465,"children":1467},{"id":1466},"outputs",[1468],{"type":55,"value":1469},"Outputs",{"type":49,"tag":58,"props":1471,"children":1472},{},[1473,1475,1480],{"type":55,"value":1474},"All outputs go to ",{"type":49,"tag":76,"props":1476,"children":1478},{"className":1477},[],[1479],{"type":55,"value":1437},{"type":55,"value":1481},":",{"type":49,"tag":111,"props":1483,"children":1484},{},[1485,1503,1552,1570,1594,1604,1615],{"type":49,"tag":115,"props":1486,"children":1487},{},[1488,1493,1495,1501],{"type":49,"tag":76,"props":1489,"children":1491},{"className":1490},[],[1492],{"type":55,"value":1429},{"type":55,"value":1494}," — initial per-image records (unless ",{"type":49,"tag":76,"props":1496,"children":1498},{"className":1497},[],[1499],{"type":55,"value":1500},"input_annotations_jsonl",{"type":55,"value":1502}," was supplied).",{"type":49,"tag":115,"props":1504,"children":1505},{},[1506,1511,1513,1519,1521,1527,1528,1533,1534,1539,1540,1545,1546,1551],{"type":49,"tag":76,"props":1507,"children":1509},{"className":1508},[],[1510],{"type":55,"value":944},{"type":55,"value":1512}," — adds ",{"type":49,"tag":76,"props":1514,"children":1516},{"className":1515},[],[1517],{"type":55,"value":1518},"regions[]",{"type":55,"value":1520}," (each with ",{"type":49,"tag":76,"props":1522,"children":1524},{"className":1523},[],[1525],{"type":55,"value":1526},"bbox",{"type":55,"value":1415},{"type":49,"tag":76,"props":1529,"children":1531},{"className":1530},[],[1532],{"type":55,"value":131},{"type":55,"value":133},{"type":49,"tag":76,"props":1535,"children":1537},{"className":1536},[],[1538],{"type":55,"value":139},{"type":55,"value":133},{"type":49,"tag":76,"props":1541,"children":1543},{"className":1542},[],[1544],{"type":55,"value":146},{"type":55,"value":133},{"type":49,"tag":76,"props":1547,"children":1549},{"className":1548},[],[1550],{"type":55,"value":153},{"type":55,"value":155},{"type":49,"tag":115,"props":1553,"children":1554},{},[1555,1561,1562,1568],{"type":49,"tag":76,"props":1556,"children":1558},{"className":1557},[],[1559],{"type":55,"value":1560},"step_1_image_caption\u002Fannotations.jsonl",{"type":55,"value":1512},{"type":49,"tag":76,"props":1563,"children":1565},{"className":1564},[],[1566],{"type":55,"value":1567},"caption",{"type":55,"value":1569}," (string).",{"type":49,"tag":115,"props":1571,"children":1572},{},[1573,1578,1579,1585,1587,1593],{"type":49,"tag":76,"props":1574,"children":1576},{"className":1575},[],[1577],{"type":55,"value":956},{"type":55,"value":1512},{"type":49,"tag":76,"props":1580,"children":1582},{"className":1581},[],[1583],{"type":55,"value":1584},"expressions[]",{"type":55,"value":1586}," (each ",{"type":49,"tag":76,"props":1588,"children":1590},{"className":1589},[],[1591],{"type":55,"value":1592},"{text, instances: [{bbox: [x1,y1,x2,y2]}]}",{"type":55,"value":155},{"type":49,"tag":115,"props":1595,"children":1596},{},[1597,1602],{"type":49,"tag":76,"props":1598,"children":1600},{"className":1599},[],[1601],{"type":55,"value":968},{"type":55,"value":1603}," — same shape as Step 2, with bboxes removed\u002Fupdated.",{"type":49,"tag":115,"props":1605,"children":1606},{},[1607,1613],{"type":49,"tag":76,"props":1608,"children":1610},{"className":1609},[],[1611],{"type":55,"value":1612},"results_dir\u002Fannotations.jsonl",{"type":55,"value":1614}," — copy of the last completed step's output.",{"type":49,"tag":115,"props":1616,"children":1617},{},[1618,1620,1625,1627,1632,1633,1638,1640,1646],{"type":55,"value":1619},"When ",{"type":49,"tag":76,"props":1621,"children":1623},{"className":1622},[],[1624],{"type":55,"value":1183},{"type":55,"value":1626}," is ",{"type":49,"tag":76,"props":1628,"children":1630},{"className":1629},[],[1631],{"type":55,"value":1217},{"type":55,"value":1254},{"type":49,"tag":76,"props":1634,"children":1636},{"className":1635},[],[1637],{"type":55,"value":1200},{"type":55,"value":1639},", each step also writes byte-compatible ",{"type":49,"tag":76,"props":1641,"children":1643},{"className":1642},[],[1644],{"type":55,"value":1645},"step_\u003CN>_*\u002Flabels\u002F\u003Cstem>.txt.stepN",{"type":55,"value":1647}," files for the original 2d-data-engine tooling.",{"type":49,"tag":64,"props":1649,"children":1651},{"id":1650},"prerequisites",[1652],{"type":55,"value":1653},"Prerequisites",{"type":49,"tag":111,"props":1655,"children":1656},{},[1657,1673,1682],{"type":49,"tag":115,"props":1658,"children":1659},{},[1660,1665,1667],{"type":49,"tag":119,"props":1661,"children":1662},{},[1663],{"type":55,"value":1664},"Container",{"type":55,"value":1666},": ",{"type":49,"tag":76,"props":1668,"children":1670},{"className":1669},[],[1671],{"type":55,"value":1672},"nvcr.io\u002Fnvidia\u002Ftao\u002Ftao-toolkit:6.26.3-pyt",{"type":49,"tag":115,"props":1674,"children":1675},{},[1676,1680],{"type":49,"tag":119,"props":1677,"children":1678},{},[1679],{"type":55,"value":348},{"type":55,"value":1681},": At least one VLM endpoint (Gemini API key or OpenAI-compatible endpoint capable of image input)",{"type":49,"tag":115,"props":1683,"children":1684},{},[1685,1690],{"type":49,"tag":119,"props":1686,"children":1687},{},[1688],{"type":55,"value":1689},"PIL \u002F Pillow",{"type":55,"value":1691},": Required to read image dimensions during seeding (already present in the TAO container)",{"type":49,"tag":1693,"props":1694,"children":1695},"style",{},[1696],{"type":55,"value":1697},"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":1699,"total":1858},[1700,1718,1736,1747,1759,1773,1786,1800,1813,1824,1838,1847],{"slug":1701,"name":1701,"fn":1702,"description":1703,"org":1704,"tags":1705,"stars":1715,"repoUrl":1716,"updatedAt":1717},"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},[1706,1709,1712],{"name":1707,"slug":1708,"type":15},"Documentation","documentation",{"name":1710,"slug":1711,"type":15},"MCP","mcp",{"name":1713,"slug":1714,"type":15},"Search","search",21777,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FNemoClaw","2026-07-20T06:00:01.461044",{"slug":1719,"name":1719,"fn":1720,"description":1721,"org":1722,"tags":1723,"stars":1733,"repoUrl":1734,"updatedAt":1735},"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},[1724,1727,1730],{"name":1725,"slug":1726,"type":15},"Containers","containers",{"name":1728,"slug":1729,"type":15},"Deployment","deployment",{"name":1731,"slug":1732,"type":15},"Python","python",17049,"https:\u002F\u002Fgithub.com\u002FNVIDIA\u002FMegatron-LM","2026-07-27T06:06:11.249662",{"slug":1737,"name":1737,"fn":1738,"description":1739,"org":1740,"tags":1741,"stars":1733,"repoUrl":1734,"updatedAt":1746},"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},[1742,1745],{"name":1743,"slug":1744,"type":15},"CI\u002FCD","ci-cd",{"name":1728,"slug":1729,"type":15},"2026-07-14T05:25:59.97109",{"slug":1748,"name":1748,"fn":1749,"description":1750,"org":1751,"tags":1752,"stars":1733,"repoUrl":1734,"updatedAt":1758},"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},[1753,1754,1755],{"name":1743,"slug":1744,"type":15},{"name":1728,"slug":1729,"type":15},{"name":1756,"slug":1757,"type":15},"GitHub","github","2026-07-27T06:06:12.278222",{"slug":1760,"name":1760,"fn":1761,"description":1762,"org":1763,"tags":1764,"stars":1733,"repoUrl":1734,"updatedAt":1772},"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},[1765,1768,1769],{"name":1766,"slug":1767,"type":15},"Debugging","debugging",{"name":1756,"slug":1757,"type":15},{"name":1770,"slug":1771,"type":15},"Triage","triage","2026-07-14T05:25:57.442089",{"slug":1774,"name":1774,"fn":1775,"description":1776,"org":1777,"tags":1778,"stars":1733,"repoUrl":1734,"updatedAt":1785},"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},[1779,1782],{"name":1780,"slug":1781,"type":15},"Best Practices","best-practices",{"name":1783,"slug":1784,"type":15},"Code Analysis","code-analysis","2026-07-14T05:25:56.18433",{"slug":1787,"name":1787,"fn":1788,"description":1789,"org":1790,"tags":1791,"stars":1733,"repoUrl":1734,"updatedAt":1799},"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},[1792,1795,1798],{"name":1793,"slug":1794,"type":15},"Machine Learning","machine-learning",{"name":1796,"slug":1797,"type":15},"Migration","migration",{"name":9,"slug":8,"type":15},"2026-07-17T06:07:11.777011",{"slug":1801,"name":1801,"fn":1802,"description":1803,"org":1804,"tags":1805,"stars":1733,"repoUrl":1734,"updatedAt":1812},"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},[1806,1809],{"name":1807,"slug":1808,"type":15},"QA","qa",{"name":1810,"slug":1811,"type":15},"Testing","testing","2026-07-14T05:25:53.673039",{"slug":1814,"name":1814,"fn":1815,"description":1816,"org":1817,"tags":1818,"stars":1733,"repoUrl":1734,"updatedAt":1823},"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},[1819,1820],{"name":1728,"slug":1729,"type":15},{"name":1821,"slug":1822,"type":15},"Infrastructure","infrastructure","2026-07-14T05:25:49.362534",{"slug":1825,"name":1825,"fn":1826,"description":1827,"org":1828,"tags":1829,"stars":1733,"repoUrl":1734,"updatedAt":1837},"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},[1830,1833,1834],{"name":1831,"slug":1832,"type":15},"Code Review","code-review",{"name":1756,"slug":1757,"type":15},{"name":1835,"slug":1836,"type":15},"Pull Requests","pull-requests","2026-07-14T05:26:01.226578",{"slug":1839,"name":1839,"fn":1840,"description":1841,"org":1842,"tags":1843,"stars":1733,"repoUrl":1734,"updatedAt":1846},"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},[1844,1845],{"name":1807,"slug":1808,"type":15},{"name":1810,"slug":1811,"type":15},"2026-07-14T05:25:54.928983",{"slug":1848,"name":1848,"fn":1849,"description":1850,"org":1851,"tags":1852,"stars":1733,"repoUrl":1734,"updatedAt":1857},"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},[1853,1856],{"name":1854,"slug":1855,"type":15},"Automation","automation",{"name":1743,"slug":1744,"type":15},"2026-07-30T05:29:03.275638",496,{"items":1860,"total":1956},[1861,1878,1888,1902,1912,1927,1942],{"slug":1862,"name":1862,"fn":1863,"description":1864,"org":1865,"tags":1866,"stars":20,"repoUrl":21,"updatedAt":1877},"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},[1867,1870,1873,1874],{"name":1868,"slug":1869,"type":15},"Data Analysis","data-analysis",{"name":1871,"slug":1872,"type":15},"Data Engineering","data-engineering",{"name":9,"slug":8,"type":15},{"name":1875,"slug":1876,"type":15},"Performance","performance","2026-07-14T05:28:43.176466",{"slug":1879,"name":1879,"fn":1880,"description":1881,"org":1882,"tags":1883,"stars":20,"repoUrl":21,"updatedAt":1887},"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},[1884,1885,1886],{"name":1728,"slug":1729,"type":15},{"name":1821,"slug":1822,"type":15},{"name":9,"slug":8,"type":15},"2026-07-14T05:29:06.667109",{"slug":1889,"name":1889,"fn":1890,"description":1891,"org":1892,"tags":1893,"stars":20,"repoUrl":21,"updatedAt":1901},"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},[1894,1897,1898],{"name":1895,"slug":1896,"type":15},"Agents","agents",{"name":9,"slug":8,"type":15},{"name":1899,"slug":1900,"type":15},"Research","research","2026-07-14T05:28:06.816956",{"slug":1903,"name":1903,"fn":1904,"description":1905,"org":1906,"tags":1907,"stars":20,"repoUrl":21,"updatedAt":1911},"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},[1908,1909,1910],{"name":1868,"slug":1869,"type":15},{"name":9,"slug":8,"type":15},{"name":1810,"slug":1811,"type":15},"2026-07-17T05:29:03.913266",{"slug":1913,"name":1913,"fn":1914,"description":1915,"org":1916,"tags":1917,"stars":20,"repoUrl":21,"updatedAt":1926},"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},[1918,1919,1922,1923],{"name":1854,"slug":1855,"type":15},{"name":1920,"slug":1921,"type":15},"Imaging","imaging",{"name":9,"slug":8,"type":15},{"name":1924,"slug":1925,"type":15},"Video","video","2026-07-17T05:28:53.905004",{"slug":1928,"name":1928,"fn":1929,"description":1930,"org":1931,"tags":1932,"stars":20,"repoUrl":21,"updatedAt":1941},"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},[1933,1934,1937,1938],{"name":1728,"slug":1729,"type":15},{"name":1935,"slug":1936,"type":15},"Docker","docker",{"name":9,"slug":8,"type":15},{"name":1939,"slug":1940,"type":15},"Operations","operations","2026-07-17T05:28:56.913999",{"slug":1943,"name":1943,"fn":1944,"description":1945,"org":1946,"tags":1947,"stars":20,"repoUrl":21,"updatedAt":1955},"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},[1948,1949,1952],{"name":9,"slug":8,"type":15},{"name":1950,"slug":1951,"type":15},"Quantum Computing","quantum-computing",{"name":1953,"slug":1954,"type":15},"Simulation","simulation","2026-07-14T05:26:58.898253",305]