README.md

metadata

license: apache-2.0
task_categories:
  - video-classification
  - zero-shot-classification
language:
  - en
tags:
  - social-signals
  - behavioral-analysis
  - multimodal
  - video
pretty_name: Wave1 Test Set - Social Signal Detection
size_categories:
  - n<1K

Wave1 Test Set - Social Signal Detection

Dataset Description

This dataset contains video clips annotated with social and affective signals for behavioral analysis tasks. Each sample includes:

• System Prompt: Instructions for behavioral analysis
• Video URL: Link to the video clip
• Multi-labels: Detected social signals from 12 categories
• Video ID: Unique identifier for the video

Social Signal Categories

The dataset includes annotations for 12 social signals:

1. Agreement: Nods, verbal affirmations, affiliative smiles, posture mirroring
2. Disagreement: Opposition, interruptions, head shakes, eye rolls
3. Engagement: Sustained orientation, eye contact, forward-leaning posture
4. Disengagement: Gaze aversion, fidgeting, low-energy speech
5. Confusion: Puzzled facial display, increased blinking, clarification questions
6. Hesitation: Long gaps before speaking, filled pauses, restarts
7. Uncertainty: Hedged answers, rising final contour, shoulder shrugs
8. Skepticism: Questioning stance, delayed responses, furrowed brows
9. Confidence: Upright posture, sustained eye contact, clear speech
10. Frustration: Tension cues, higher pitch, sighs, complaints
11. Interest: Attentional engagement, widened eyes, faster speech
12. Stress: Self-touch, repetitive fidgeting, rapid blinking, vocal instability

Data Format

The dataset is provided as a JSONL file where each line contains:

{
  "messages": [
    {
      "content": "System prompt with task instructions...",
      "role": "system"
    },
    {
      "content": "<video>",
      "role": "user"
    },
    {
      "content": "",
      "role": "assistant"
    }
  ],
  "videos": ["https://..."],
  "multilabels": ["Signal1", "Signal2", ...],
  "video_id": "unique_video_identifier.mp4"
}

Usage

from datasets import load_dataset

# Load dataset
dataset = load_dataset("Interhuman/wave1-test-set", split="train")

# Or load the raw JSONL file
import json

with open("configs_wave1-test-set-dataset.jsonl", "r") as f:
    data = [json.loads(line) for line in f]

# Access first sample
sample = data[0]
print(sample['video_id'])
print(sample['multilabels'])
print(sample['videos'][0])

Dataset Statistics

• Total Samples: 50
• Multi-label Classification: Each video can have multiple social signals
• Signal Distribution: Varies across samples

Citation

If you use this dataset, please cite:

@dataset{interhuman_wave1_test,
  title={Wave1 Test Set - Social Signal Detection},
  author={Interhuman},
  year={2024},
  publisher={Hugging Face}
}

License

Apache 2.0