sop-data-augmentation

To convert annotated videos into QA data for VLM fine-tuning, configure a YAML augmentation pipeline to orchestrate asynchronous jobs through a SOP data-generation service, producing LLaVA-format training datasets.

For SOP monitoring datasets, you can generate seven distinct QA formats: BCQ, sequential MCQ, golden GQA, LLM-expanded GQA, dynamic MCQ, dynamic shuffling, and extra-negative or cross-SOP negative training data.

You do not strictly need a local vLLM server to generate grounded question-answer data; you can alternatively use NVIDIA NIM access for LLM-expanded GQA generation and thinking-mode model configuration.

To troubleshoot a failed data augmentation job for vision-language models, validate your annotated dataset structure, inspect progress logs, verify generated QA directories, and resolve service or LLM errors using the provided guidance.

Yes, you can use Docker to run the SOP data augmentation pipeline, as it is a required dependency alongside vLLM for orchestrating the service and generating structured question-answer training data.

Hard negative mining in SOP video data augmentation generates cross-SOP negative and extra-negative training examples to challenge Vision-Language Models, ensuring robust monitoring workflow performance across diverse scenarios.