Subnet 24

Omega

OMEGA Labs

OMEGA Labs subnet is building the world’s largest decentralized multimodal dataset to advance AGI research

SN24 : Omega Labs

SN24 : Omega Labs

SubnetDescriptionCategoryCompany
SN24 : Omega LabsMultimodal datasetsData Pipeline
Omega Labs

SN1 SN2 SN3 SN4 SN5 SN6 SN7 SN8 SN9 SN10 SN11 SN12 SN13 SN14 SN15 SN16 SN17 SN18 SN19 SN20 SN21 SN22 SN23 SN24 SN25 SN26 SN27 SN28 SN29 SN30 SN31 SN32 SN33 SN34 SN35 SN36 SN37 SN38 SN39 SN40 SN41 SN42 SN43 SN44 SN45 SN46 SN47 SN48 SN49 SN50 SN51 SN52 SN53 SN54 SN55 SN56 SN57 SN58 SN59 SN60 SN61 SN62 SN63 SN64

The OMEGA Labs Bittensor subnet, is an innovative initiative focused on creating the world’s largest decentralized multimodal dataset to advance research and development in Artificial General Intelligence (AGI). Their mission is to democratize access to an extensive and diverse dataset that encompasses human knowledge and creativity, empowering researchers and developers to push the boundaries of AGI.

Harnessing the Bittensor network and a global community of miners and validators, they are constructing a dataset that exceeds the scale and diversity of existing resources. Featuring over 1 million hours of footage and 30 million+ 2-minute video clips, the OMEGA Labs dataset will facilitate the development of robust AGI models and drive transformation across multiple industries.

Key Features

  • Unmatched Scale and Variety: Over 1 million hours of footage and 30 million video clips spanning 50+ scenarios and 15,000+ action phrases.
  • Latent Representations: Utilizing cutting-edge models to translate video components into a unified latent space for efficient processing.
  • Incentivized Data Collection: Rewarding miners for contributing high-quality, diverse, and innovative videos through a decentralized network.
  • Empowering Digital Agents: Facilitating the development of intelligent agents capable of navigating complex workflows and supporting users across platforms.
  • Immersive Gaming Experiences: Enabling the creation of realistic gaming environments with rich physics and interactive elements.

Miner

Conducts searches on YouTube and retrieves up to 8 videos per query. Specifies a clip range (up to 2 minutes) and provides a description (catch) including video title, tags, and description. Obtains ImageBind embeddings for video, audio, and caption components. Returns video ID, caption, ImageBind embeddings (video, audio, caption), and start and end times for clips (up to 2 minutes).

Validator

Randomly selects one video from those submitted by miners for validation. Calculates ImageBind embeddings for all modalities (video, audio, caption) of the selected video. Compares embeddings to ensure consistency with miner submissions. If validated, assumes all eight videos from the miner are valid. Scores videos based on relevance, novelty, and detail richness:

Detail Richness: Determines similarity between text and video embeddings. Collects 1024 validated video entries and submits them as a concatenated file to Hugging Face. Adjusts file accumulation limits if a miner submits too frequently. Submits remaining validated entries in case of API shutdowns.

Relevance: Uses cosine similarity between topic embeddings and each of the eight videos.

Novelty: Calculates 1 – similarity to the closest video in the Pinecone index.

Ben-Zion Benkhin – Founder and CEO

Ben founded WOMBO in 2020, aiming to simplify cutting-edge technology for everyday use.

Salman Shahid – Machine Learning Engineer

Salman, fascinated by autonomous AI since a young age, viewed AI as a tool for exploring groundbreaking concepts.

Parshant Loungani – Founder and Head of AI

Parshant, with a physics background, transitioned into AI due to its potential for innovation, leading to the creation of the successful WOMBO app.