This repository collects the benchmark sources used for task-aware visual-token sensitivity experiments in StreamingVLM. The goal is to cover four task groups:

Task group Dataset keys Purpose

Coarse object presence pope, repope, hpope Tests whether low token budgets preserve semantic object existence.

General scene understanding vqav2, gqa Tests ordinary VQA scene/action/relation understanding.

Fine-grained visual evidence… See the full description on the dataset page: https://huggingface.co/datasets/kfkas/streamingvlm-task-aware-vqa-suite.