N-back_1

This repository is a private transfer bundle for building multi-video N-back sequences.

It contains:

• raw source videos
• pair manifests
• pre-cut segment packages

Main Idea

Each pair contains two highly similar videos from the same dataset.

The intended similarity structure is:

• within the same pair: highly similar
• across different pairs from the same dataset: generally similar
• across different datasets: relatively dissimilar

This makes the package suitable for constructing both:

• more similar source pools
• less similar source pools

In practice, a simple workflow is:

1. Choose one pair as the anchor pair.
2. If you want another similar source pair, choose a different pair from the same dataset.
3. If you want a less similar source pair, choose a pair from a different dataset.
4. Use either the raw videos or the pre-cut segment packages below.

Repository Layout

Raw Pair Package

• pairs/nback_1_pairs_20_per_dataset.json Combined manifest with 20 selected pairs from each dataset.
• pairs/<dataset>_pairs.json Per-dataset manifests for easier browsing.
• pairs/manifest_summary.json Summary counts.
• videos/<dataset>/<pair_id>/video_1.mp4
• videos/<dataset>/<pair_id>/video_2.mp4

The raw package is the right choice if you want to cut or sample the videos yourself.

Pre-cut Packages

Pre-cut segment bundles are stored under folders such as:

• precut_subsets/...
• precut_full_20pairs_per_dataset_seg10/...

Each pre-cut package has its own README and manifest. These are the right choice if you want to start directly from equal-count video segments.

Recommended Usage

If You Want Raw Videos

1. Open pairs/nback_1_pairs_20_per_dataset.json.
2. Choose one pair as the anchor pair.
3. Choose another pair from the same dataset if you want a more similar source pool.
4. Choose another pair from a different dataset if you want a less similar source pool.
5. Read the corresponding raw files from: videos/<dataset>/<pair_id>/video_1.mp4 and videos/<dataset>/<pair_id>/video_2.mp4

This is the recommended path if you want full control over:

• how many clips to cut from each video
• clip duration
• ordering strategy
• prompt construction

If You Want Pre-cut 10-Way Segments

1. Open the README inside the corresponding pre-cut folder.
2. Use the package manifest there to locate segment files.
3. Each source video is split into: seg00 to seg09
4. Use the manifest time fields as the planned time windows for downstream construction.

This is the recommended path if you want to move quickly and do not want to cut videos on your own server.

Example Sequence Construction

Suppose you want to build one source pool for a multi-video N-back sequence.

1. Pick one anchor pair from pairs/nback_1_pairs_20_per_dataset.json.
2. Use the two videos in that pair as a highly similar base.
3. Add another pair from the same dataset if you want more semantically related videos.
4. Or add a pair from a different dataset if you want a less similar source pool.
5. If you use a pre-cut package, read the corresponding seg00 to seg09 files directly.
6. If you use raw videos, cut them with your own strategy.

Pair Manifest Fields

The pair JSON files include:

• pair_id
• dataset
• selection_rank_within_dataset
• video_1_id
• video_2_id
• original_video_1_path
• original_video_2_path
• repo_video_1_path
• repo_video_2_path

repo_video_*_path is the path inside this Hugging Face repository.

The per-pair JSON also keeps the selection metadata used during pair construction, which can help if you want to reuse the same sampling logic.

Datasets Covered

• CrossVid-CC
• CrossVid-NC
• HourVideo
• InfiniBench-TVQA
• LVBench
• Video-MME-L

Notes

• The videos come from the original benchmark datasets used in VideoMemoryEval.
• Some pre-cut packages are produced with fast ffmpeg stream-copy segmentation.
• In those packages, downstream users should rely on the manifest time windows rather than assuming frame-perfect physical cut points.
• If you only need one urgent subset, check the precut_subsets/ folder first.
• If you want the complete 20-pairs-per-dataset transfer bundle with 10-way cuts, use precut_full_20pairs_per_dataset_seg10/.