arkimjh/werewolf2.0

Werewolf 2.0 Benchmark

Description

Werewolf 2.0 is a video-based social reasoning benchmark for evaluating gaze and gesture understanding in multi-person scenarios. Each video clip is paired with a multiple-choice QA testing fine-grained social cue comprehension.

Dataset Statistics

	Count
Video clips	1,196
QA pairs (MCQ)	1,196
Categories	16 (T1-T6, G1-G6, C1-C4)
Answer choices	4 (A/B/C/D)
Video FPS	2 fps

16-Category QA Taxonomy

Cat	Name	Difficulty
T1	Gaze Target Identification	Easy
T2	Gaze Event Classification	Easy
T3	Mutual Gaze Recognition	Medium
T4	Gaze Following	Medium
T5	Temporal Gaze Reasoning	Hard
T6	Group Attention Dynamics	Hard
G1	Gesture Recognition	Easy
G2	Gesture Type Classification	Easy
G3	Gesture Temporal Reasoning	Medium
G4	Reciprocal Gesture Patterns	Medium
G5	Gesture Frequency	Hard
G6	Gesture Sequence Chains	Hard
C1	Gaze-Gesture Temporal Alignment	Medium
C2	Gaze Response to Gesture	Medium
C3	Eye Contact During Interaction	Hard
C4	Cross-Modal Person Dynamics	Hard

Dataset Structure

werewolf2.0/
├── tars/
│   ├── werewolf2.0_videos_part001.tar   → video/*.mp4
│   └── werewolf2.0_jsons_part001.tar    → json/*.json
└── README.md

Each JSON file contains:

{
  "video_name": "153_1_001",
  "clip_file": "153_1_001_qa0000.mp4",
  "clip_offset": 5.5,
  "clip_duration": 5.5,
  "qa": {
    "category": "T1",
    "difficulty": "easy",
    "format": "mcq",
    "question": "Between 1.5s and 3.5s, whose gaze does Person 3 follow?",
    "options": ["A) Person 0", "B) Person 2", "C) Person 4", "D) No one"],
    "answer": "A",
    "answer_text": "Person 0"
  }
}

Usage

from huggingface_hub import snapshot_download
import tarfile
from pathlib import Path

# Download
snapshot_download(
    repo_id="arkimjh/werewolf2.0",
    repo_type="dataset",
    local_dir="./werewolf2.0"
)

# Extract
for tar_file in Path("./werewolf2.0/tars").glob("*.tar"):
    with tarfile.open(tar_file) as tf:
        tf.extractall("./werewolf2.0/")

Citation

TBD