Datasets:

VideoMemBench
/

data

@@ -22,41 +22,279 @@ configs:
 # VideoMemBench
-<p>
-  <img src="https://raw.githubusercontent.com/Oliiveralien/VideoMemBench/main/assets/readme-banner.svg" width="100%" style="display: block; width: 100%; height: auto;" alt="VideoMemBench: How Well Do MLLMs Remember in Video Conversations?" />
 </p>
-VideoMemBench is a benchmark of multi-turn video dialogues in which later questions can only be answered
-by remembering what earlier turns established. Each dialogue is grounded in one video. A dependent turn
-refers back to earlier turns either through a scene or entity they established (episodic dependency) or
-through their answers (answer dependency).
-[Paper](https://arxiv.org/abs/2610.40131v1) · [Project](https://videomembench.github.io) · [GitHub](https://github.com/Oliiveralien/VideoMemBench)
 ## Data format
 ```json
-{"video_id": "videomembench-1", "video_name": "videomembench-1.mp4", "video_path": "videos/videomembench-1.mp4",
- "dialogue": [
-   {"turn_id": "T1", "type": "open_structured", "dependency": null,
-    "question": "...", "options": null, "answer": "..."},
-   {"turn_id": "T2", "type": "multi_choice_question",
-    "dependency": {"depends_on": ["T1"], "dep_type": {"T1": "episodic"}},
-    "question": "...", "options": {"A": "...", "B": "...", "C": "...", "D": "..."}, "answer": "C"}
- ]}
 ```
-| Field | Description |
-|---|---|
-| `turn_id` | Turn index within the dialogue (`T1`, `T2`, ...) |
-| `type` | `multi_choice_question`, `true_or_false_question` or `open_structured` |
-| `dependency` | `null` for an independent turn; otherwise `depends_on` lists the parent turns and `dep_type` maps each parent to `episodic` or `answer` |
-| `question` | Question text |
-| `options` | Four to eight lettered options for multiple-choice turns; `null` otherwise |
-| `answer` | Option letter, `True` / `False`, or a free-form reference answer |
-An `answer` dependency means the turn needs the parent turn's correct answer. An `episodic` dependency
-means the turn returns to a scene or entity established in the parent turn.
 ## License and terms of use

 # VideoMemBench
+<p align="center">
+  <img src="https://raw.githubusercontent.com/Oliiveralien/VideoMemBench/main/assets/readme-banner.svg" width="100%" alt="VideoMemBench: How Well Do MLLMs Remember in Video Conversations?" />
 </p>
+<p align="center">
+  <a href="https://videomembench.github.io/"><img src="https://img.shields.io/badge/%F0%9F%8D%8E%20Project-VideoMemBench-e58ab5?style=flat-square" alt="🍎 Project: VideoMemBench" height="26" /></a>
+  <a href="https://arxiv.org/abs/2610.40131"><img src="https://img.shields.io/badge/cs.CV-arXiv%3A2610.40131-b31b1b?style=flat-square&logo=arxiv&logoColor=white" alt="cs.CV: arXiv:2610.40131" height="26" /></a>
+  <a href="https://github.com/Oliiveralien/VideoMemBench"><img src="https://img.shields.io/badge/GitHub-VideoMemBench-2855d9?style=flat-square&logo=github&logoColor=white" alt="GitHub: VideoMemBench" height="26" /></a>
+  <a href="https://videomembench.github.io/#results"><img src="https://img.shields.io/badge/%F0%9F%8F%86%20Leaderboard-Rank-e6b000?style=flat-square" alt="🏆 Leaderboard: Rank" height="26" /></a>
+</p>
+**Zhengdong Hu\*, Chao Wang\*, Kaituo Feng, Huanjin Yao, Mingfei Lu, Hehe Fan, Yi Yang†**
+\* Equal contribution. † Corresponding author.
+## Benchmark
+VideoMemBench evaluates **in-context video memory**: the ability to retrieve and reuse information established earlier in a multi-turn conversation about a video. Each sample pairs one video with an ordered dialogue, reference answers, and typed dependencies between turns.
+- **Episodic dependency:** a later question revisits a visual episode, entity, or event established in an earlier turn.
+- **Answer dependency:** a later question builds on an earlier answer, requiring information carried through the conversation.
+Dependencies are checked for necessity and non-leakage, and answers must be visually grounded. A planner, generator, and judger construct and check dialogues, followed by human expert review. Videos are drawn from Video-MME, VSI-Bench, Video-MMMU, LongVideoBench, EgoLife, and YouTube. See the [paper](https://arxiv.org/html/2610.40131v1) for construction and collection details.
+<p align="center">
+  <img src="https://raw.githubusercontent.com/Oliiveralien/VideoMemBench/main/assets/hero.svg" width="100%" alt="VideoMemBench benchmark overview: multi-turn video conversations with episodic and answer dependencies" />
+</p>
+| Property | Value |
+| --- | ---: |
+| Videos / dialogues | 326 / 326 |
+| QA turns | 2,476 |
+| Independent / dependent turns | 1,089 / 1,387 |
+| Episodic / Answer dependency edges | 1,200 / 594 |
+| Turns with multiple parents | 348 |
+| Open / multiple-choice / true-or-false questions | 1,709 / 603 / 164 |
+| Turns per dialogue | 4–21; mean 7.6 |
+| Semantic domains / broad categories | 31 / 4 |
+The four broad categories are Lifestyle & Entertainment (130 videos), Knowledge & Education (80), Art & Literature (66), and Sports & Competition (50). A turn may have multiple parents and both dependency types. Each annotation record maps to one video in `videos/`.
+## Leaderboard
+The main evaluation uses **128 fixed, timestamped frames per video** and **sequential self-rollout**: each model answers in dialogue order and retains its own earlier responses. Scores are percentages; higher is better. Rows follow the official ranking by **Overall**, the mean of TurnAcc, EpiDepAcc, and AnsDepAcc.
+Results below are copied from the [official project leaderboard](https://videomembench.github.io/#results), checked on **2026-11-01**.
+### Main results · 128 frames
+| Rank | Model | TurnAcc | EpiDepAcc | AnsDepAcc | Overall ↑ |
+| ---: | --- | ---: | ---: | ---: | ---: |
+| 1 | Claude Opus 5 | 52.3 | 49.1 | 35.7 | 45.7 |
+| 2 | Gemini 3.8 Flash | 50.4 | 46.6 | 30.3 | 42.4 |
+| 3 | Kimi K3 | 50.5 | 49.0 | 27.1 | 42.2 |
+| 4 | GPT-5.6-sol | 50.0 | 47.8 | 27.1 | 41.7 |
+| 5 | Gemini 3.5 Flash | 49.6 | 45.5 | 29.7 | 41.6 |
+| 6 | GPT-5.5 | 48.0 | 46.5 | 27.0 | 40.5 |
+| 7 | Claude Sonnet 5 | 47.8 | 45.3 | 27.3 | 40.2 |
+| 8 | Gemini 3.1 Pro Preview | 45.7 | 44.8 | 27.0 | 39.1 |
+| 9 | GPT-5.4 | 44.4 | 42.6 | 21.9 | 36.3 |
+| 10 | Qwen3-VL-235B-A22B | 43.1 | 41.6 | 18.6 | 34.4 |
+| 11 | Qwen3.5-27B | 42.2 | 39.8 | 20.9 | 34.3 |
+| 12 | Qwen3.5-397B-A17B | 42.3 | 40.1 | 20.3 | 34.2 |
+| 13 | Qwen3.5-122B-A10B | 42.4 | 39.5 | 19.9 | 33.9 |
+| 14 | Qwen3.5-35B-A3B | 38.2 | 36.3 | 18.0 | 30.8 |
+| 15 | Qwen3-VL-32B | 39.3 | 35.9 | 17.2 | 30.8 |
+| 16 | Qwen3-VL-30B-A3B | 36.4 | 35.1 | 15.8 | 29.1 |
+| 17 | Qwen3.5-9B | 34.6 | 33.4 | 12.9 | 27.0 |
+| 18 | Qwen3-VL-8B | 34.8 | 32.1 | 12.3 | 26.4 |
+| 19 | Qwen3.5-4B | 33.8 | 32.6 | 12.5 | 26.3 |
+| 20 | Qwen3-VL-4B | 32.4 | 29.4 | 12.9 | 24.9 |
+| 21 | Qwen2.5-VL-32B | 31.7 | 32.6 | 9.6 | 24.6 |
+| 22 | Qwen2.5-VL-7B | 26.2 | 24.9 | 8.4 | 19.8 |
+| 23 | Keye-VL-1.5-8B | 25.4 | 24.5 | 8.2 | 19.4 |
+| 24 | Qwen2.5-VL-3B | 23.5 | 24.4 | 5.7 | 17.9 |
+| 25 | Qwen3.5-0.8B | 21.6 | 21.3 | 5.7 | 16.2 |
+- **TurnAcc:** accuracy across all turns.
+- **EpiDepAcc:** accuracy on turns with at least one Episodic Dependency, independent of parent correctness.
+- **AnsDepAcc:** credit requires both the current response and every recursive Answer Dependency ancestor to be correct; Episodic Dependency edges are excluded from ancestor traversal.
+### Frame-count comparison
+The [official frame-count comparison](https://videomembench.github.io/#frame-results) evaluates 13 open-weight models with 64 or 32 uniformly sampled frames, using the same self-rollout protocol and scoring rules.
+<details>
+<summary><strong>64 frames · 13 models</strong></summary>
+| Rank | Model | TurnAcc | EpiDepAcc | AnsDepAcc | Overall ↑ |
+| ---: | --- | ---: | ---: | ---: | ---: |
+| 1 | Qwen3-VL-235B-A22B | 40.3 | 39.3 | 15.6 | 31.7 |
+| 2 | Qwen3.5-122B-A10B | 37.8 | 36.5 | 15.2 | 29.8 |
+| 3 | Qwen3.5-27B | 37.9 | 36.7 | 14.8 | 29.8 |
+| 4 | Qwen3-VL-32B | 37.8 | 36.1 | 14.5 | 29.5 |
+| 5 | Qwen3.5-35B-A3B | 37.0 | 36.7 | 13.5 | 29.1 |
+| 6 | Qwen3-VL-30B-A3B | 33.6 | 32.9 | 11.5 | 26.0 |
+| 7 | Qwen3-VL-8B | 32.7 | 32.0 | 11.9 | 25.5 |
+| 8 | Qwen3.5-9B | 33.4 | 33.2 | 9.6 | 25.4 |
+| 9 | Qwen3-VL-4B | 31.3 | 30.3 | 12.1 | 24.5 |
+| 10 | Keye-VL-1.5-8B | 24.5 | 24.9 | 6.3 | 18.5 |
+| 11 | Qwen2.5-VL-7B | 24.1 | 22.7 | 6.4 | 17.7 |
+| 12 | Qwen2.5-VL-3B | 22.5 | 22.5 | 5.1 | 16.7 |
+| 13 | Qwen3.5-0.8B | 21.0 | 21.1 | 4.7 | 15.6 |
+</details>
+<details>
+<summary><strong>32 frames · 13 models</strong></summary>
+| Rank | Model | TurnAcc | EpiDepAcc | AnsDepAcc | Overall ↑ |
+| ---: | --- | ---: | ---: | ---: | ---: |
+| 1 | Qwen3-VL-235B-A22B | 36.3 | 37.8 | 14.5 | 29.5 |
+| 2 | Qwen3.5-122B-A10B | 35.0 | 35.2 | 10.7 | 27.0 |
+| 3 | Qwen3.5-35B-A3B | 34.5 | 34.1 | 12.3 | 27.0 |
+| 4 | Qwen3.5-27B | 34.9 | 34.2 | 11.7 | 26.9 |
+| 5 | Qwen3-VL-32B | 34.1 | 32.6 | 13.1 | 26.6 |
+| 6 | Qwen3-VL-30B-A3B | 32.1 | 32.5 | 8.0 | 24.2 |
+| 7 | Qwen3.5-9B | 29.8 | 30.4 | 8.6 | 22.9 |
+| 8 | Qwen3-VL-8B | 29.4 | 29.4 | 9.2 | 22.7 |
+| 9 | Qwen3-VL-4B | 29.0 | 29.0 | 8.8 | 22.3 |
+| 10 | Qwen2.5-VL-7B | 23.2 | 23.7 | 4.9 | 17.3 |
+| 11 | Keye-VL-1.5-8B | 23.3 | 23.1 | 5.1 | 17.2 |
+| 12 | Qwen2.5-VL-3B | 22.6 | 23.4 | 4.7 | 16.9 |
+| 13 | Qwen3.5-0.8B | 20.1 | 21.8 | 4.7 | 15.5 |
+</details>
+## Download and load
+The repository contains one evaluation split, `test`:
+```text
+annotations.jsonl        # 326 video-level dialogue records
+videos/
+  videomembench-1.mp4
+  ...
+  videomembench-326.mp4
+```
+### Load annotations
+```bash
+pip install huggingface_hub
+```
+```python
+import json
+from huggingface_hub import hf_hub_download
+annotation_path = hf_hub_download(
+    repo_id="VideoMemBench/data",
+    repo_type="dataset",
+    filename="annotations.jsonl",
+)
+with open(annotation_path, encoding="utf-8") as source:
+    dialogues = [json.loads(line) for line in source if line.strip()]
+print(len(dialogues))  # 326
+print(dialogues[0]["video_path"])
+print(dialogues[0]["dialogue"][0]["question"])
+```
+### Download videos for evaluation
+The full video collection is approximately **146.5 GB** (136.4 GiB). The following downloads the annotations and videos together:
+```python
+from huggingface_hub import snapshot_download
+data_dir = snapshot_download(
+    repo_id="VideoMemBench/data",
+    repo_type="dataset",
+    local_dir="data",
+    allow_patterns=["annotations.jsonl", "videos/*.mp4"],
+)
+```
+Downloads use your local Hugging Face authentication when access is required.
 ## Data format
+`annotations.jsonl` is UTF-8 JSONL: each line contains one video and its ordered dialogue.
+| Field | Description |
+| --- | --- |
+| `video_id` | Unique video identifier, such as `videomembench-1`. |
+| `video_name` | Video filename, such as `videomembench-1.mp4`. |
+| `video_path` | Path relative to the repository root, such as `videos/videomembench-1.mp4`. |
+| `dialogue` | Questions in their original conversation order. |
+| `dialogue[].turn_id` | Turn identifier within the dialogue (`T1`, `T2`, ...). |
+| `dialogue[].type` | `open_structured`, `multi_choice_question`, or `true_or_false_question`. |
+| `dialogue[].question` | Question text. |
+| `dialogue[].options` | Lettered options for multiple-choice questions; `null` otherwise. |
+| `dialogue[].answer` | Option key, a True/False reference, or a free-form reference answer. |
+| `dialogue[].dependency` | `null` for an independent turn; otherwise the parent IDs and dependency types. |
+A dependency object contains `depends_on`, a list of earlier turn IDs, and `dep_type`, which maps each parent ID to `episodic` or `answer`. Preserve turn order during evaluation. Reference answers and dependency labels are evaluation metadata and must not be supplied to the answering model.
+<details>
+<summary><strong>Example record structure</strong></summary>
+The placeholder text below illustrates the structure and is not a benchmark question. JSONL stores each record on one line; it is expanded here for readability.
 ```json
+{
+  "video_id": "videomembench-1",
+  "video_name": "videomembench-1.mp4",
+  "video_path": "videos/videomembench-1.mp4",
+  "dialogue": [
+    {
+      "turn_id": "T1",
+      "type": "open_structured",
+      "dependency": null,
+      "question": "<independent question>",
+      "options": null,
+      "answer": "<reference answer>"
+    },
+    {
+      "turn_id": "T2",
+      "type": "multi_choice_question",
+      "dependency": {
+        "depends_on": ["T1"],
+        "dep_type": {"T1": "episodic"}
+      },
+      "question": "<dependent question>",
+      "options": {"A": "<option A>", "B": "<option B>", "C": "<option C>", "D": "<option D>"},
+      "answer": "C"
+    }
+  ]
+}
 ```
+</details>
+## Evaluation
+Use the [official evaluation code](https://github.com/Oliiveralien/VideoMemBench) with Python 3.11 or newer:
+```bash
+git clone https://github.com/Oliiveralien/VideoMemBench.git
+cd VideoMemBench
+pip install -e .
+```
+Place the downloaded files under `data/`. The Hub records include `video_path` for locating video files. The evaluation code resolves files from `video_name` and `--video-root`, and its strict validator does not accept the extra `video_path` field. Create an evaluation copy as follows:
+```python
+import json
+from pathlib import Path
+data_dir = Path("data")
+with (data_dir / "annotations.jsonl").open(encoding="utf-8") as source, \
+     (data_dir / "annotations.eval.jsonl").open("w", encoding="utf-8") as target:
+    for line in source:
+        if not line.strip():
+            continue
+        record = json.loads(line)
+        record.pop("video_path", None)
+        target.write(json.dumps(record, ensure_ascii=False) + "\n")
+```
+Validate the evaluation copy and prepare the shared 128-frame cache:
+```bash
+python -m videomembench.data data/annotations.eval.jsonl
+python -m videomembench.video \
+  --annotations data/annotations.eval.jsonl \
+  --video-root data/videos \
+  --frame-cache cache/f128
+```
+Set the annotation path to `data/annotations.eval.jsonl`, configure the model and judge endpoints in [`configs/evaluation.toml`](https://github.com/Oliiveralien/VideoMemBench/blob/main/configs/evaluation.toml), and follow the repository's inference and scoring instructions.
+The main evaluation uses 128 uniformly sampled, timestamped frames with a maximum side length of 512 pixels. Models answer turns sequentially while retaining their own full response history. Reference answers, dependencies, audio, and subtitles are not provided to the answering model. Open answers are graded by GPT-5.4 mini. See the [evaluation protocol](https://github.com/Oliiveralien/VideoMemBench/blob/main/docs/PROTOCOL.md) for complete details.
 ## License and terms of use