CUE-Mem Benchmark Dataset

This repository contains the publicly released benchmark data used by CUE-Mem. The repository is organized so that every path stored in a JSON or JSONL file is relative to the dataset repository root.

Repository layout

CUE-Mem-Benchmark/
├── README.md
├── LICENSE
├── data/
│   ├── dialog/base/                  # Per-profile dialogue and QA JSON files
│   ├── event/images/                 # Event images
│   ├── event/voice_mixed/            # Event audio files
│   ├── qa/pref_images/               # Preference-question images
│   ├── qa/rec_images/                # Recognition-question images
│   ├── qa/entity_images/             # Entity-question images
│   └── profile/generated_portraits/  # Generated profile portraits
└── metadata/
    ├── dialogue_turns.jsonl
    ├── questions.jsonl
    ├── media_manifest.jsonl
    └── checksums.sha256

The audio directory can be populated in a later upload. Whenever files are added or replaced, regenerate metadata/media_manifest.jsonl and metadata/checksums.sha256.

Path convention

Media paths embedded in the JSON and JSONL files use POSIX separators and are relative to the dataset repository root. For example:

data/event/images/pid_0000_task_0-12-0.png
data/qa/pref_images/0-FoodAndDrink-0_C.png

Metadata files

• dialogue_turns.jsonl contains one record per dialogue turn.
• questions.jsonl contains one record per human-annotated question.
• media_manifest.jsonl indexes the media files, their type, format, and size.
• checksums.sha256 contains SHA-256 checksums for all files under data/.

Basic usage

Run the following from a local clone of this dataset repository:

from pathlib import Path
import json

dataset_root = Path(".")
with (dataset_root / "metadata/dialogue_turns.jsonl").open(encoding="utf-8") as f:
    turn = json.loads(next(f))

for relative_path in turn.get("input_image", []):
    image_path = dataset_root / relative_path
    print(image_path)

To verify the downloaded data:

sha256sum -c metadata/checksums.sha256

License

See LICENSE for the applicable license terms.