README.md

---
pretty_name: SuperMemory-VQA
task_categories:
- visual-question-answering
tags:
- egocentric-video
- imu
- streaming-vqa
- benchmark-manifest
---

# SuperMemory-VQA

Viewer-ready benchmark manifest for `kfkas/supermemory-vqa-imu-benchmark`. It contains benchmark
annotations and logical local asset hints, not redistributed source media.

## Status

- Base dataset: `SuperMemory-VQA Aria stream`
- Rows: `4,853`
- Rows with local RGB+IMU: `4,818`
- Answer-key status: `public=4853`
- Existing repository refreshed with the common schema.

## Example video

A short, recompressed benchmark-corresponding clip is included for a quick
visual check. It is not the complete source recording.

<video controls width="640" preload="metadata"><source src="https://huggingface.co/datasets/kfkas/supermemory-vqa-imu-benchmark/resolve/main/examples/sample.mp4" type="video/mp4"></video>

[Open the example video](https://huggingface.co/datasets/kfkas/supermemory-vqa-imu-benchmark/resolve/main/examples/sample.mp4)

## Common schema

All rows use the same columns: `benchmark`, `base_dataset`, `question_id`,
`category`, `question`, `choices`, `answer_index`, `answer_text`, `source_ids`,
`query_time`, `evidence_time`, `clip_start_s`, `clip_end_s`, `stream_count`,
`video_asset_hint`, `imu_asset_hint`, `local_video_available`,
`local_imu_available`, `usable_for_imu_benchmark`, `coverage_status`,
`answer_key_status`, `metadata_json`, and `streams`.

`streams` uses a consistent nested shape for source ID, modality, sensor role,
time bounds, asset hints, local availability, and join kind. Dataset-specific
fields are encoded in `metadata_json` so the viewer schema remains stable.

## Load

```python
from datasets import load_dataset

dataset = load_dataset("kfkas/supermemory-vqa-imu-benchmark", split="benchmark")
usable = dataset.filter(lambda row: row["usable_for_imu_benchmark"])
```

No full raw MP4, VRS, ZIP, CSV, or NPZ files are included. Only the short
example clip is provided. Download complete source media from the upstream
terms and use the relative hints under `sensor_vqa/raw/`.

## Sources

- https://huggingface.co/datasets/OSU-AIoT-MLSys-Lab/SuperMemory-VQA