icomgpu/ppi

Educational Video Patterns Dataset

Dataset Description

This dataset contains processed video segments with audio transcriptions and metadata. The dataset focuses on educational patterns extracted from video recordings, with automatic language detection supporting both English (primary) and Russian languages.

Dataset Summary

This multimodal dataset includes:

• Video segments (.mp4) - cropped video clips extracted from source recordings
• Images (.jpg) - key frames extracted from video segments
• Audio (.wav) - extracted audio tracks (16kHz, mono)
• Transcriptions (.txt) - text transcriptions of audio content
• Metadata (metadata.csv) - comprehensive information about each sample

Supported Tasks

• Audio-to-text transcription: Automatic speech recognition with multi-language support
• Video classification: Pattern recognition in educational contexts
• Image classification: Key frame analysis

Languages

The dataset primarily contains English audio with some Russian content. Language detection is performed automatically during transcription using Whisper models.

Dataset Structure

Data Fields

The dataset contains the following fields (in order):

• image - Image (PIL Image object)

  • Key frame extracted from the video segment
  • Format: JPEG

• audio - Audio file (Audio object)

  • Extracted audio track from video
  • Format: WAV, 16kHz, mono

• pattern_name - Pattern name (string)

  • Name/identifier of the educational pattern

• description - Pattern description (string)

  • Optional description of the pattern (if available)

• transcription - Transcription text (string)

  • Text transcription from metadata.csv
  • Language automatically detected (English/Russian)

• start_time - Start time (float64)

  • Start time of the segment in the source video (seconds)

• end_time - End time (float64)

  • End time of the segment in the source video (seconds)

• video - Video file path (string)

  • Path to the video file (file is available when loading the dataset)

Data Splits

The dataset contains a single split:

• train: All samples

Dataset Creation

Source Data

The dataset was created from video recordings annotated using Label Studio, with the following processing pipeline:

1. Video Processing (main.py):

   • Parsed JSON annotations from Label Studio
   • Created cropped video segments based on temporal markers
   • Extracted key frames from video segments

2. Dataset Preparation (prepare_dataset.py):

   • Copied video and image files
   • Extracted audio tracks to WAV format (16kHz, mono)
   • Performed audio transcription using Whisper (with automatic language detection)
   • Created metadata CSV with all sample information

3. Quality Control:

   • Manual filtering to remove low-quality or irrelevant samples
   • Frame-by-frame review and deletion of unwanted segments

4. Upload (upload_to_hf.py):

   • Dataset uploaded to Hugging Face Hub
   • Dataset card generated automatically

Annotations

• Annotation process: Manual annotation using Label Studio
• Annotation guidelines: Educational patterns identified and marked with temporal boundaries
• Who annotated: Dataset creators

Personal and Sensitive Information

This dataset contains educational video content. No personal or sensitive information is expected, but users should review the content before use.

Considerations for Using the Data

Social Impact of Dataset

This dataset is intended for educational and research purposes, focusing on teaching pattern recognition and analysis.

Discussion of Biases

The dataset may reflect biases present in the source educational materials. Users should be aware of potential language, cultural, or educational biases.

Other Known Limitations

• Limited sample size (n<1K)
• Primary language is English with some Russian content
• Manual filtering may introduce subjective quality criteria

Additional Information

Dataset Curators

The dataset was curated by the project maintainers.

Licensing Information

MIT License - see LICENSE file for details.

Citation Information

If you use this dataset, please cite:

@dataset{educational_video_patterns,
  title={Educational Video Patterns Dataset},
  author={Dataset Authors},
  year={2024},
  license={MIT}
}

Contributions

Contributions and improvements to the dataset are welcome.

Usage

Loading the Dataset

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("username/dataset-name")

# Access data
print(dataset['train'][0])

# Access specific fields
sample = dataset['train'][0]
image = sample['image']
audio = sample['audio']
transcription = sample['transcription']

Example Use Cases

• Speech Recognition Training: Train models on English/Russian educational content
• Video Analysis: Analyze teaching patterns and techniques
• Multimodal Learning: Combine video, audio, and text for educational research
• Pattern Recognition: Identify and classify educational patterns

Preprocessing

The dataset is preprocessed and ready to use. Audio files are normalized to 16kHz mono, and video segments are cropped to relevant time ranges.

Dataset Statistics

• Total samples: 10
• Dataset size: 3.67 MB (3,665,949 bytes)
• Download size: 3.61 MB (3,613,992 bytes)
• Primary language: English
• Secondary language: Russian
• Audio format: WAV, 16kHz, mono
• Video format: MP4
• Image format: JPEG
• Split: train (10 examples)

Updates and Versions

• Version 1.0: Initial release with manual filtering and quality control