ProgramerSalar/Rita_Black_white_mix

📚 Rita: Class 8 Math Video Dataset (Raw)

Primary Training Data for Zulense Z1

This dataset contains raw video footage of Class 8 Mathematics lectures (Indian Curriculum/NCERT), categorized by teaching medium (Blackboard vs. Whiteboard) and teacher presence.

This data is used to train the Zulense Imagination Engine to understand the difference between:

1. Teacher Dynamics: How a teacher moves and gestures.
2. Content Logic: How equations and diagrams appear on a board over time.

📂 Folder Structure & Contents

The dataset is partitioned into ZIP archives based on the visual setting.

Archive Name	Content Description	Duration Strategy
long_video_black_board.zip	Traditional Classroom: Full-length lectures on a classic chalk blackboard. Teacher is visible (pose/face visible). Used for learning teacher-board interaction.	Long Form
long_video_black_board_without_teacher.zip	POV / Focus Mode: Lectures on a blackboard where the teacher is behind the camera or out of frame. Focus is 100% on the chalk writing and diagrams.	Long Form
long_video_white_board.zip	Modern Classroom: Full-length lectures on a marker whiteboard. Teacher is visible. High contrast environment.	Long Form
long_video_white_board_without_teacher.zip	POV / Focus Mode: Whiteboard lectures with no teacher visible. Used to train the model on pure text/diagram generation without occlusion.	Long Form
long_video_clip.zip	Short Samples: Curated 5-10 second clips of whiteboard teaching. Useful for testing temporal consistency and overfitting on short sequences.	Short Clips
long_video_mix_board.zip	Synthetic / Mixed Media: A collection of outliers including animated avatars, digital screen recordings, and graphic-heavy explanations. Used for regularization.	Mixed

🎯 Training Objective

We use this segregated data to perform Concept Disentanglement in our Diffusion Models:

• Teacher-Free folders (_without_teacher) are used to teach the VAE how to reconstruct text and geometry with high fidelity.
• Teacher-Visible folders are used to train the DiT (Diffusion Transformer) on human movement and pointing gestures.
• Black/White Separation ensures the model can generalize across different contrast domains (Chalk on Black vs. Marker on White).

📝 Usage

from datasets import load_dataset

# Load the dataset (conceptual example)
dataset = load_dataset("ProgramerSalar/Rita_Black_white_mix", split="train")

# The data is stored in ZIP files, requiring extraction before processing.