README.md

---
task_categories:
  - visual-question-answering
language:
  - en
tags:
  - video-understanding
  - eventmemagent
configs:
  - config_name: training
    data_files:
      - split: train
        path: train/eventmemagent_train.parquet
  - config_name: ovobench
    data_files:
      - split: test
        path: eval/ovobench/eventmemagent_ovobench.parquet
  - config_name: streamingbench
    data_files:
      - split: test
        path: eval/streamingbench/eventmemagent_streamingbench.parquet
---

# EventMemAgent

Prepared training and evaluation inputs for **EventMemAgent: Hierarchical
Event-Centric Memory for Online Video Understanding with Adaptive Tool Use**
(ECCV 2026).

[Paper](https://arxiv.org/abs/2602.15329) ·
[Code](https://github.com/lingcco/EventMemAgent) ·
[Model](https://huggingface.co/lingcco/EventMemAgent-8B)

## Contents

| Configuration | Split | Rows | Description |
| --- | --- | ---: | --- |
| training | train | 10,000 | MovieChat samples with VideoMarathon annotations |
| ovobench | test | 3,035 | OVO-Bench inputs with prepared event memory |
| streamingbench | test | 2,500 | StreamingBench inputs with prepared event memory |

The evaluation folders also contain `ovo_bench_new.json` and `questions_real.json`,
the reference annotations needed by the scoring scripts. Generated predictions,
scoring outputs and experiment logs are not included. No separate training
validation split is included.

## Memory construction

Videos are sampled at 1 FPS. Short-term memory contains at most 32 frames;
event segmentation uses minimum event length 8 and content threshold 0.2.
Frozen Qwen3-VL-4B-Instruct generates event captions and
Qwen3-Embedding-0.6B generates event embeddings.

Parquet records retain the original nested fields, including `prompt`,
`question`, `ground_truth`, `agent_name`, `reward_model`, `data_source`,
`short_term_memory`, `long_term_memory` and `extra_info.tools_kwargs`. Images
are embedded in memory records. Original video paths identify samples and do
not need to exist locally for precomputed-memory evaluation.

## Loading

```python
from datasets import load_dataset

train = load_dataset("lingcco/EventMemAgent", "training", split="train", streaming=True)
ovo = load_dataset("lingcco/EventMemAgent", "ovobench", split="test", streaming=True)
streamingbench = load_dataset("lingcco/EventMemAgent", "streamingbench", split="test", streaming=True)
```

Alternatively, download the parquet files and reference JSON files and use the
accompanying code repository's evaluation and scoring scripts. Keep evaluation
splits out of policy optimization and checkpoint selection.

## Sources and usage terms

These processed inputs are derived from MovieChat, VideoMarathon, OVO-Bench
and StreamingBench. Original video frames and annotations remain subject to
their respective upstream licenses and usage terms; this repository does not
grant additional rights to third-party content. The accompanying code's Apache
2.0 license does not apply automatically to the underlying datasets.