---
license: mit
task_categories:
- video-classification
- text-to-video
tags:
- large-scale
- ncert
- math
- education
- raw-video
- zulense
pretty_name: "Rita: 350GB Raw Educational Video Corpus"
size_categories:
- 100G<n<1T
---

# 📚 Rita: Large-Scale Raw Educational Video Corpus (348 GB)

**The Foundation Layer for the [Zulense Imagination Engine](https://zulense.com)**

This repository contains the **unfiltered, raw video ingestion** used for the pre-training of Zulense's educational models. It consists of approximately **348 GB** of Class 8 Mathematics lectures (aligned with the Indian NCERT curriculum), aggregated from diverse open educational sources.

## 💾 Dataset Statistics

* **Total Size:** ~348 GB
* **Format:** Raw Video Files (`.mp4`, `.mkv`)
* **Subject:** Mathematics (Algebra, Geometry, Data Handling, Mensuration)
* **Curriculum:** Class 8 (Indian Standard)
* **Diversity:** Includes various lighting conditions, teaching styles (blackboard, whiteboard, digital pad), and audio qualities.

## 🏗️ The Data Pipeline

This dataset represents **Layer 0 (Raw Ingestion)** of our training pipeline. It is not meant for direct fine-tuning but for **Self-Supervised Video Pre-training**.

The pipeline flows as follows:

1.  **Layer 0 (This Dataset):** **Massive Raw Corpus (348GB).** Used to teach the model general video concepts (how a hand moves, how chalk appears, how a classroom looks).
2.  **Layer 1 (Filtering):** We filter this data into "clean" subsets (removing intros, outros, and non-teaching segments).
3.  **Layer 2 (Classification):** Data is sorted into Blackboard/Whiteboard categories (See: [Rita_Black_white_mix](https://huggingface.co/datasets/ProgramerSalar/Rita_Black_white_mix)).
4.  **Layer 3 (Latent Encoding):** Data is compressed via VAE for the final training run.

## ⚠️ Storage Warning & Usage

**This is a massive dataset.** Do not attempt to clone this repository unless you have significant disk space and bandwidth.

### Recommended Usage (Streaming)

We recommend using the Hugging Face `streaming=True` feature or downloading specific shards if you are exploring the data.

```python
from datasets import load_dataset

# Stream the dataset without downloading 350GB locally
dataset = load_dataset("ProgramerSalar/Rita", split="train", streaming=True)

print("Iterating through video samples...")
for sample in dataset.take(5):
    print(f"Video Path: {sample['path']}")