SLVMBench: Skill Learning from Video Memory

[NeurIPS 2026 Datasets & Benchmarks Track]

Official Hugging Face Dataset repository for SLVMBench, a comprehensive benchmark designed to evaluate whether Video Large Language Models (video-LLMs) can acquire procedural skills from long video memory and apply them to real-time, ongoing tasks under heavy distractor noise.

---

📂 Repository Structure

To optimize download efficiency and bandwidth, this repository contains all original video clips (tutorials, distractors, and target tasks) and full JSON annotations. The 2-hour long-context concatenated video streams can be easily generated locally using our provided lightweight script.

SLVMBench/
├── videos/                   # Folder containing all raw video clips (.mp4)
│   ├── tutorial_xxx.mp4
│   ├── distractor_yyy.mp4
│   └── target_zzz.mp4
├── SLVMBench.json            # Complete QA annotations and video stream metadata
├── concat_videos.py    # Python script to concatenate raw videos into streaming inputs
├── README.md                 # Dataset documentation
└── .gitattributes

---

🛠️ How to Generate Long-Context Video Streams (2-Hour Concatenated Streams)

Because concatenated 2-hour streaming videos are memory-intensive, we provide build_merged_videos.py to seamlessly build the full evaluation streams locally on demand.

Quick Start:

1. Clone the Repository:

   git clone https://huggingface.co/datasets/SLVMBench/SLVMBench
   cd SLVMBench
   

2. Install Dependencies:

   pip install moviepy tqdm
   

3. Run Video Concatenation Script:

   python build_merged_videos.py --annotation SLVMBench.json --video_dir videos/ --output_dir merged_videos/
   

   This script will automatically read the distractor sequences and cutoff timestamps specified in SLVMBench.json to generate the concatenated continuous video streams ready for evaluation.

---

📊 Dataset Format (JSON Schema)

The SLVMBench.json file contains 2,261 human-verified Q&A pairs. Each entry follows this structure:

{
  "id": "sample_001",
  "domain": "Cooking",
  "orig_video_path": "videos/tutorial_01.mp4",      // Tutorial video teaching the skill
  "video_path": "videos/target_01.mp4",           // Target task execution video
  "distractors": [                                // Sequence of unrelated distractor videos
    "videos/distractor_01.mp4",
    "videos/distractor_02.mp4"
  ],
  "duration_sec": 3600.0,                          // Total merged video stream duration
  "question": {
    "start": 12.0,                                 // Key action start time
    "end": 15.0,                                   // Decision point cutoff timestamp (t_c)
    "question": "What is the next correct step to perform?",
    "options": [
      "Option A ...",
      "Option B ...",
      "Option C ...",
      "Option D ...",
      "Option E ..."
    ],
    "correct_choice": "A"
  }
}

---

🔬 Evaluation Paradigms

SLVMBench evaluates video-LLMs across three standardized paradigms:

1. w/o Tutorial (Baseline): Only the target video truncated at $t_c$ is provided (measures common-sense priors).
2. w/ Tutorial (Immediate): The tutorial video followed immediately by the truncated target video.
3. w/ Tutorial (Long): The tutorial video, followed by a sequence of distractor videos (up to 2 hours), followed by the truncated target video.

---

📜 Licensing & Citation

All metadata and annotations are released under the Apache 2.0 License. Video clips are collected from public YouTube sources for academic evaluation purposes under Fair Use guidelines.