README.md

---
license: cc-by-4.0
task_categories:
- visual-question-answering
- video-text-to-text
language:
- en
tags:
- video
- video-question-answering
- long-video-understanding
- entity-tracking
- hallucination
- benchmark
- evaluation
size_categories:
- 1K<n<10K
configs:
- config_name: default
  data_files:
  - split: test
    path: narrativetrack_qa.json
---

# NarrativeTrack

This is the official release of the dataset for the paper
**[NarrativeTrack: Evaluating Entity-Centric Reasoning for Narrative Understanding](https://arxiv.org/abs/2601.01095)**.

📄 Paper: [https://arxiv.org/abs/2601.01095](https://arxiv.org/abs/2601.01095) &nbsp;|&nbsp;
🤗 Hugging Face Paper: [https://huggingface.co/papers/2601.01095](https://huggingface.co/papers/2601.01095) &nbsp;|&nbsp;
💻 Project page: [https://github.com/apple/ml-NarrativeTrack](https://github.com/apple/ml-NarrativeTrack)

**NarrativeTrack** is an evaluation benchmark for measuring how well video–language models
**track entities across long, narrative videos** — and how prone they are to **hallucinating**
identities, actions, outfits, and scenes when an entity appears, disappears, and reappears
over time.

Each example presents a short video clip of a target entity together with a question that
probes whether the model can correctly relate the entity's appearance at one point in the
video to another point (e.g. *"Is the person shown at the end the same person who was shown
behind fences earlier?"*). Distractors are drawn from real co-occurring entities as well as
synthetically perturbed attributes, making the benchmark a stress test for cross-time
entity consistency.

This is an **evaluation-only** benchmark: it provides a single **`test`** split (no training
split). The clips are sourced from [AVA](https://research.google.com/ava/),
[Video-MME](https://video-mme.github.io/), and
[LVBench](https://lvbench.github.io/).

## Dataset Statistics

- **Examples:** 1,006 QA pairs
- **Videos:** 406 unique entity clips
- **Sources:** Video-MME (509), AVA (282), LVBench (215)
- **Question types:** `binary` (478), `mc` / multiple-choice (446), `ordering` (82)
- **Tracking types:** `appear` (271), `reappear` (586), `disappear` (149)
- **Dimensions:** `entity_existence`, `entity_ambiguity`, `entity_action_changes`,
  `entity_outfit_changes`, `entity_scene_changes`

## Data Fields

Each record in `narrativetrack_qa.json` contains:

| Field | Type | Description |
|-------|------|-------------|
| `id` | int | Unique question id |
| `video_path` | str | Path to the entity clip, **relative to the dataset root** (e.g. `videos/AVA/1j20qq1JyX4/entity2_3/appear/1320_1350.mp4`). After extracting `videos.tar` this path resolves directly. |
| `question` | str | The question shown to the model (answer options are included inline for `mc` and `ordering`). |
| `answer` | str | Ground-truth answer. `Yes`/`No` for `binary`, an option letter (`a`–`d`) for `mc`, and a comma-separated option ordering (e.g. `a,c,b`) for `ordering`. |
| `question_type` | str | One of `binary`, `mc`, `ordering`. |
| `dimension` | str | The reasoning dimension being probed (existence, ambiguity, action / outfit / scene changes). |
| `track_type` | str | Entity tracking event: `appear`, `reappear`, `disappear`. |
| `template_type` | str | Temporal framing of the question (e.g. `later_to_start`, `start_to_later`, `agnostic`). |
| `template` | str | The slot-filled template the question was generated from. |
| `entity_id` | str | Identifier of the target entity within the source video. |
| `distractor_type` | str | Source of distractor options: `real`, `synthetic`, or `mix`. |

## Download Videos

```py
from huggingface_hub import hf_hub_download
import tarfile
import shutil

file_path = hf_hub_download(
    repo_id="hjha/NarrativeTrack",
    filename="videos.tar",
    repo_type="dataset",
)

print("Downloaded to:", file_path)
shutil.copy(file_path, "./videos.tar")

with tarfile.open("videos.tar", "r") as tar:
    tar.extractall(path=".")
# This creates ./videos/... matching the `video_path` field in each record.
```

## Load Dataset

```py
from datasets import load_dataset

# Evaluation benchmark: only a `test` split is available.
ds = load_dataset("hjha/NarrativeTrack")["test"]

example = ds[0]
print(example["question"])
print("answer:", example["answer"])
print("video:", example["video_path"])  # e.g. videos/AVA/.../1320_1350.mp4
```

Or, as a PyTorch `Dataset` that resolves `video_path` against the extracted videos:

```py
import os
from datasets import load_dataset
from torch.utils.data import Dataset


class NarrativeTrackDataset(Dataset):
    def __init__(self, video_root="."):
        # Point `video_root` at the directory where you extracted videos.tar
        # (the one that contains the `videos/` folder).
        self.video_root = video_root
        self.data = load_dataset("hjha/NarrativeTrack")["test"]

    def __len__(self):
        return len(self.data)

    def __getitem__(self, idx):
        sample = dict(self.data[idx])
        sample["video_path"] = os.path.join(self.video_root, sample["video_path"])
        assert os.path.exists(sample["video_path"]), \
            f"Video not found: {sample['video_path']}"
        return sample


test_ds = NarrativeTrackDataset(video_root=".")
example = test_ds[0]
```

## Files

- `narrativetrack_qa.json` — the 1,006 QA records (test split).
- `videos.tar` — the 406 entity clips, stored under `videos/<source>/...` with paths matching
  the `video_path` field.

## Citation

If you use NarrativeTrack in your research, please cite:

```bibtex
@article{narrativetrack2026,
  title   = {NarrativeTrack: Evaluating Entity-Centric Reasoning for Narrative Understanding},
  journal = {arXiv preprint arXiv:2601.01095},
  year    = {2026},
  url     = {https://arxiv.org/abs/2601.01095}
}
```