| --- |
| license: cc-by-4.0 |
| task_categories: |
| - visual-question-answering |
| - image-text-to-text |
| language: |
| - en |
| tags: |
| - streaming-vlm |
| - task-aware-resolution |
| - visual-token-sensitivity |
| - pope |
| - textvqa |
| - docvqa |
| - vqav2 |
| - gqa |
| - mmbench |
| pretty_name: StreamingVLM Task-Aware VQA Suite |
| --- |
| |
| # StreamingVLM Task-Aware VQA Suite |
|
|
| 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 | `gqa_attribute`, `mmbench` | Tests non-OCR attribute, relation, part, and perception sensitivity. | |
| | Textual fine-grained evidence | `textvqa`, `docvqa` | Tests OCR/document detail sensitivity. | |
|
|
| ## Files |
|
|
| - `source_registry.jsonl`: canonical source registry for every benchmark. |
| - `manifests/*.jsonl.gz`: locally materialized full manifests that were already |
| used in StreamingVLM experiments. |
|
|
| The local manifests include: |
|
|
| | Manifest | Rows | |
| |---|---:| |
| | `manifests/pope_coco_manifest_n9000.jsonl.gz` | 9,000 | |
| | `manifests/repope_train_manifest_n8185.jsonl.gz` | 8,185 | |
| | `manifests/hpope_train_manifest_n9904.jsonl.gz` | 9,904 | |
| | `manifests/textvqa_validation_manifest_n5000.jsonl.gz` | 5,000 | |
| | `manifests/docvqa_validation_manifest_n5349.jsonl.gz` | 5,349 | |
|
|
| For `VQAv2`, `GQA`, and `MMBench`, this repository stores the full source |
| references rather than rehosting the full image payloads. Those datasets are |
| large, and the experiment runner should load them from the original Hugging Face |
| source repositories listed in `source_registry.jsonl`. |
|
|
| ## 500-Sample Experiment Protocol |
|
|
| This repository stores full dataset sources or full local manifests. Individual |
| experiments should select 500 examples per dataset at runtime: |
|
|
| ```text |
| full source / full manifest |
| -> deterministic 500-row selection |
| -> resolution or token-compression sweep |
| ``` |
|
|
| Use the same selection seed across all visual-token budgets and model variants |
| so that performance differences reflect token allocation rather than sample |
| composition. |
|
|
| ## Recommended Dataset Mapping |
|
|
| ```text |
| Coarse/object: |
| POPE, RePOPE, H-POPE |
| |
| General VQA: |
| VQAv2, GQA |
| |
| Fine-grained object/attribute: |
| GQA attribute subset, MMBench |
| |
| Text/OCR: |
| TextVQA, DocVQA |
| ``` |
|
|
| ## Notes |
|
|
| The manifests contain absolute local paths from the original experiment machine |
| when images were materialized locally. Use the `source_repo`, `source_split`, |
| `image_id`, and question/answer fields as the stable dataset identity. Image |
| payloads should be resolved from the corresponding original dataset source when |
| running on a different machine. |
|
|
