paidf-auto-labeling

Confirm critical inputs first: data path, output path, VLM/LLM endpoints and model names, model cache, and GPU ids. Then run a shipped example cookbook with --container-dry-run to prove the stack, plan your scenario, adapt the closest cookbook, author prompts, and dry-run before executing.

Use the shipped cookbook at cookbooks/video_data_augmentation/configs/pipeline_video.yaml through workflow-runner:main. Always run with --container-dry-run first, pass pre-flight checks for media, SAM3 weights, endpoints, and GPUs, then obtain explicit approval before the real Docker run.

Choose RF-DETR when COCO classes and stable track IDs are sufficient, such as traffic or person inventory. Choose SAM3 when target objects are not COCO classes, when masks or contours are required, or when a VLM should reason over an annotated ID video.

Reasoning-capable models such as Gemini 3 Flash spend part of the token budget on internal thinking, so a low max_tokens cap truncates outputs. Raise max_tokens to a generous value like 32768 on the visual_qa and reasoning LLM substages for reasoning models.

No. Committed cookbooks must use placeholder paths and bare environment variable names like NVIDIA_API_KEY, which forwards the host value into containers at runtime. Never write NAME=value secrets into YAML, logs, or dry-run output.

Use the pipeline-migration path when adopting an external annotation or dataset-generation repository into PAIDF. It classifies each source stage as cookbook-only reuse, an extension of an existing service, or a new generic capability service with a stable sidecar contract.