Subnet 33
ReadyAI
Afterparty
ReadyAI fosters a decentralized effort to build an extensive dialogue dataset
SN33 : ReadyAI
| Subnet | Description | Category | Company |
|---|---|---|---|
| SN33 : ReadyAI | Data structuring | Data Pipeline | Afterparty |
ReadyAI provides creators and businesses with cost-effective, time-saving tools to build the most accurate AI applications for their audience. Built on the Bittensor network, they’ve transformed data tagging and structuring for text data, significantly improving AI accuracy through metadata and synthetic data generation for vector databases. These databases give you control over your data while leveraging top-tier models like OpenAI, Anthropic, and Meta.
ReadyAI offers more precise metadata tagging at a lower cost than competitors like Scale AI, which rely on expensive, less reliable human labor. Their use of Bittensor’s decentralized validation ensures higher quality control over metadata, and they implement a privacy-preserving system for processing both public and proprietary data. As they continue to evolve, ReadyAI is committed to protecting data privacy, fostering collaboration in AI research, and advancing their dataset to meet the growing demands of conversational AI, helping create seamless, human-like interactions for emotional support, personalized learning, and beyond.
Afterparty, the team behind ReadyAI, aims to enable creators to fully own and monetize their AI personas and creations, highlighting the need for creators to have true ownership of their AI platforms. Afterparty recognizes a lack of quality in open-source foundational models for conversational natural language fluency.
The ReadyAI subnet is designed to offer a low-cost, resource-efficient data structuring and semantic tagging pipeline for both individuals and businesses. A prime example of their data processing power is the creation of an annotated, persona-labeled dialogue dataset for conversational AI development—The ReadyAI. This dataset provides crucial resources for the open-source community to make significant strides in developing natural, engaging conversational AI through further training, finetuning, and system integration.
ReadyAI leverages Bittensor’s decentralized infrastructure to incentivize a global network of miners and validators to contribute and verify high-quality conversational data. Their innovative “fractal data mining” approach enhances data integrity by cross-referencing miner submissions with ground truth, rewarding accurate and valuable contributions.
Key Features
- Indexing and tagging billions of conversations from diverse sources such as YouTube and podcasts
- Utilizing fractal data mining and conversation windows to ensure efficient and privacy-preserving processing
- Generating synthetic participant profiles using conversation metadata
- Implementing an algorithm to evaluate conversation quality based on relevance, engagement, novelty, coherence, and fluency
- Providing an open-source dataset for training and refining conversational AI models
- Establishing an incentivized mining and validation system to ensure data contribution and maintain integrity
ReadyAI utilizes Bittensor’s infrastructure to annotate conversational data.
Benefits
- Addresses the current lack of personalization in conversational AI models
- Facilitates natural and engaging conversations tailored to individual contexts and preferences
- Provides a comprehensive and annotated dataset for advancing conversational AI
- Promotes contributions and innovations within the open-source community
- Ensures data integrity through validation and scoring mechanisms
System Design
- Data stores: Serve as the primary source of truth for conversation windows, participant profiles, and vector databases
- Validator roles: Involve pulling data, generating overview metadata for grounded conversations, creating windows, and scoring submissions. Validators have the authority to decide what data to allow or block, ensuring personal information remains confidential. Maintaining control over data protection is fundamental to the design.
- Miner roles: Focus on processing conversation windows, contributing metadata, and applying tags. Specific metadata tags are extracted to assess their similarity to those in the overall conversation, focusing on uniqueness to provide a deeper understanding of the conversation’s context. The conversational window safeguards against miners creating unauthorized databases with personal data. Miners have the opportunity to enhance their performance by utilizing system prompting, high-quality chaining, and fine-tuning models to improve metadata extraction. The flexibility for miners to fine-tune their models based on their sources allows for endless opportunities in the mining process.
- Data flow: Includes establishing ground truth, creating windows, submitting miner contributions, scoring, and validation
Rewards and Incentives
An algorithm is in place for assessing conversation quality, though it is not yet used for miner rewards.
Miners are rewarded for contributing accurate and valuable metadata. The “conversation Windows” concept in Subnet divides conversations into overlapping segments, allowing miners to analyze each segment and extract detailed metadata. Miners receive a portion of the conversation to analyze, providing a more granular level of metadata tagging compared to the overall conversation tags.
Rewards are distributed in a balanced manner to encourage high-quality submissions. Miners’ scoring is based on matching tags with the overall conversation and generating unique tags close to the semantic neighborhood of the conversation. Validators compare miners’ tags within the same conversation window to ensure relevance and uniqueness in the metadata tagging process.
Cross-referencing and vector embedding analysis are used to ensure data integrity.
The Afterparty team consists of individuals with backgrounds in AI, technology, and creator marketing, showcasing a diverse range of expertise to disrupt the creator economy.
David Fields – Co-Founder
Eytan Elbaz – Co-Founder
Robert Graham – Co-Founder and Chief Talent Officer
Dan Rahmel – Co-Founder
John Van Liere – Product Lead
Matthew Dusette – Product Operations Manager