Sarvesh-369/Low-Frequency-Trap

The Low-Frequency Trap Benchmark Dataset

Official dataset repository for "The Low-Frequency Trap: Video–Language Models Fail at Simple Event Bookkeeping".

---

📌 Dataset Overview

The Low-Frequency Trap Benchmark evaluates Video–Language Models (VLMs) on fine-grained visual event bookkeeping across controlled parametric variations of Event Load (N) and Event Frequency (F).

Rather than evaluating models solely on final aggregate integer counts, this benchmark pairs every video clip with an executable ground-truth trace containing precise event timestamps (t_i), event types (e_i), state transitions (s_i⁻ → s_i⁺), and running cumulative counts (c_i).

• Total Video Samples: 2,190 controlled synthetic video clips (730 per domain across 3 visual domains).
• Video Duration: Fixed 24.0 seconds per clip rendered in high definition.
• Event Load Range (N): N ∈ {0, 1, 2, 3, 4, 5, 6, 8, 10, 12}
• Event Frequency Range (F): F ∈ {0.5, 1.0, 1.5, 2.0, 2.5, 3.0, 3.5, 4.0} Hz

---

🎥 Task Domains

The dataset covers three distinct visual domains testing complementary physical and state dynamics:

1. bounce_ball: Wall collision bookkeeping of a moving ball within a bounded 2D arena (transient physical contact events).
2. blinking: Periodic visual pulse opacities of a central object (transient visual flash events).
3. state_machine: Categorical node highlights and color shifts across an indicator panel (persistent state transition events).

---

📁 Repository Structure

Dataset/
├── manifest.jsonl        # Master catalog containing metadata for all 2,190 samples
├── load_dataset.py       # Helper script to load and inspect manifest entries
├── videos/               # Rendered MP4 clips organized by domain
│   ├── bounce_ball/
│   ├── blinking/
│   └── state_machine/
├── gt/                   # Ground-truth event details per sample
├── questions/            # Domain task questions
├── reasoning_traces/     # Diagnostic trace templates
├── traces/               # Executable ground-truth trace logs
├── .gitattributes        # Hugging Face Git LFS configuration
└── README.md             # Dataset Card

---

📄 Manifest Field Schema

Each record in manifest.jsonl follows this JSON structure:

{
  "sample_id": "bounce_N4_F1.5_seed2",
  "task_name": "bounce_ball",
  "control_parameter_name": "N",
  "control_parameter_value": 4,
  "frequency_hz": 1.5,
  "seed": 2,
  "video_path": "videos/bounce_ball/bounce_N4_F1.5_seed2.mp4",
  "question": "How many times did the ball contact the walls?",
  "exact_answer": "4",
  "duration": 24.0,
  "checksum": "b216271ca66ace42bfd7c9a0...",
  "ground_truth_reasoning": "**Question:** How many times did the ball contact the walls?\n\nLet me analyze...",
  "executable_trace": { "steps": [...], "final_count": 4, "events": [...] }
}

---

🚀 Quickstart & Usage

1. Cloning via Git LFS

git clone https://huggingface.co/datasets/Low-Frequency-Trap/Low-Frequency-Trap
cd Low-Frequency-Trap

2. Loading Manifest in Python

import json

manifest = []
with open("manifest.jsonl", "r", encoding="utf-8") as f:
    for line in f:
        manifest.append(json.loads(line))

print(f"Loaded {len(manifest)} evaluation samples.")
sample = manifest[0]
print("Sample Video Path:", sample["video_path"])
print("Question:", sample["question"])
print("Target Answer:", sample["exact_answer"])

---

📄 Citation

If you use this dataset in your research, please cite:

@misc{baskar2026lowfrequencytrapvideo,
      title={The Low Frequency Trap: Video Language Models Fail at Simple Event Bookkeeping}, 
      author={Sarvesh Baskar and Zikui Cai and Shayan Shabihi and Anirudh Satheesh and Muhammad R. Islam and Udari Madhushani Sehwag and Tom Goldstein and Furong Huang},
      year={2026},
      eprint={2608.06361},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2608.06361}, 
}