Opportunities and Vision
With the explosive demand for large language models (LLMs), the market is in urgent need of reliable models and high-quality data. However, most of the current progress is concentrated in a few closed-source, general models, while industries like finance and healthcare have an even more pressing demand for customized, high-performance models. DecideAI was born to build an open, transparent, and secure AI ecological infrastructure that protects user privacy while rewarding data contributors and promoting collaborative innovation through on-chain technology and a native token incentive mechanism. The project believes that the future focus of LLMs will shift from scaling to data quality and collaboration models. DecideAI will prioritize the development of a model training system that is of excellent quality, participatory by users, and verifiable.
The DecideAI ecology consists of three main modules: Decide Protocol (training protocol), Decide ID (identity verification), and Decide Cortex (model sharing). The Decide Protocol is responsible for collaborative annotation and model training, Decide ID ensures contributor identity and data quality, and Decide Cortex provides a platform for sharing pre-trained models and data. Users and developers can combine these three modules to jointly create high-quality dedicated LLMs. Meanwhile, the DecideAI ecology rewards data creators, annotators, and developers with the native token DCD and relies on a robust reward system to incentivize long-term participation and suppress malicious behavior, forming a sustainable collaborative environment.
Decide Protocol: Training and Data Collaboration Platform
Decide Protocol is an end-to-end platform for model training in specialized fields, providing full-process services from data collection, annotation, model training to continuous iteration for enterprise users lacking internal model development resources. This process relies on deep collaboration between artificial intelligence and human experts: after the model generates preliminary results, professional annotators review, correct, and provide feedback to iteratively improve the model. This 'Reinforcement Learning + Human Feedback (RLHF)' approach can produce outputs that align more with human expectations, balancing cost-effectiveness and data quality. Compared to fully automated fine-tuning methods, RLHF has advantages in customization and accuracy: it allows for precise adjustments to model responses based on specific goals and improves model accuracy and safety in key areas through continuous enhanced training.
In terms of technical implementation, the Decide Protocol first introduced the DeBERTa v3 model to assess the uncertainty of generated responses. This improved BERT architecture, combined with heteroscedastic uncertainty metrics, can determine the reliability of answers and guide whether more expert review is needed. The model training adopts an RLHF process based on the TRLX framework, combined with the advanced Proximal Policy Optimization (PPO) algorithm. The PPO algorithm maximizes expected rewards while maintaining policy stability, continuously receiving signals from the DeBERTa reward model and adjusting the LLM's behavior accordingly until the expected quality is met. This process ensures training quality while reducing the possibility of contributor manipulation of the system.
To improve annotation and data quality, the Decide Protocol also introduces various evaluation and optimization mechanisms: data Shapley values are used to measure each data point's contribution to model performance and are weighted based on contribution size during reward model retraining; impact functions estimate the influence of individual annotations on model predictions; cross-validation assesses the generalization capability of the reward model on different data subsets. Additionally, the system assigns suitable tasks to the most experienced annotators based on contributors' professional qualifications and historical performance to enhance participation enthusiasm and matching. All these technological improvements collectively optimize the annotation results, making the final trained model more accurate and reliable.
In the annotation process, the Decide Protocol relies on the Decide ID module to verify contributor qualifications. Only users verified through the on-chain Proof of Personhood (PoP) can qualify for participating in training data annotation. Decide ID issues a unique on-chain identity (Principal ID) for each contributor and ensures the authenticity and uniqueness of each participant through biometrics and other verifications. Thus, each training data point can be traced back to verified real contributors, ensuring the reliability of the data source. The entire Decide Protocol not only utilizes artificial intelligence technology to enhance efficiency but also ensures high quality and traceability of data through human expert feedback and strict identity verification.
Decide ID: Identity Verification and Data Quality Assurance
Decide ID provides a unique identity verification system, centered around the Proof of Personhood (PoP) method, aimed at eliminating the participation of bots or fake accounts. Every user who wishes to contribute data and annotations must generate a unique on-chain identity ID (Principal ID) in the system and submit multiple proof materials (such as biometric information, educational certificates, etc.) for verification. PoP ensures that only real, unique humans participate in the contribution process, thereby enhancing the credibility of data and annotation results from the source. This mechanism also features flexibility and portability: the verification process can be customized according to needs (e.g., verification steps and difficulty), and the generated identity ID can be reused by other applications that support PoP.
During the verification process, Decide ID employs 'Zero-Knowledge Proof (ZK Proof) technology' to protect user privacy. When users submit identity or qualification information, the system generates an encrypted ZK proof instead of directly exposing sensitive data to third parties. The verifier only needs to confirm the user's identity information through the ZK proof without seeing the original data, thus ensuring privacy security while meeting compliance requirements (such as GDPR). In other words, users retain complete control over their personal data, and no verification process will leak their private information.
In the DecideAI ecology, Decide ID closely collaborates with other modules as the foundation for quality assurance. For example, in the Decide Protocol, all contributors involved in generating and annotating training data must pass the Decide ID verification to ensure their professional background matches the task. This way, each data point is linked to a verified expert contributor, and any data anomalies or quality issues can be traced back. This on-chain identity verification and privacy protection mechanism not only ensures high quality in the annotation process but also provides a trust foundation for multi-party collaboration within the ecology.
Decide Cortex: Model and Data Sharing Platform
Decide Cortex is an open platform dedicated to inclusive AI technology, providing developers, enterprises, and researchers with access to pre-trained models and high-quality datasets. Users can either directly purchase or subscribe to models and data offered by the DecideAI community or call upon these resources via APIs. A variety of pre-trained models will be launched on the platform, covering two main categories: one is general foundational models suitable for various natural language processing tasks like text generation, summarization, and question answering; the other is customized models tailored for specific industries or tasks, excelling in scenarios such as healthcare and finance. All models and datasets released on Cortex will be continuously iterated and updated to ensure that users can always use the latest and optimal model performance.
Through Decide Cortex, users can not only access existing models but also fine-tune and retrain models according to their own needs. The platform supports the monitoring and management of models after they go live, enabling enterprises to customize exclusive models that align with their private data and business logic. In this open community model, both individual developers and institutions can contribute their own models and data, enriching ecological resources together. Overall, Decide Cortex provides a one-stop service for all ecological participants, lowering the threshold for building dedicated AI systems and sharing the high-quality model results produced by Decide Protocol with a broader user base.
Core Technology Highlights
The DecideAI architecture integrates various cutting-edge technologies: Proof of Personhood (PoP) ensures that contributors are unique, real users; 'Zero-Knowledge Proof (ZK Proof)' is used to verify identities and credentials without disclosing privacy data; DeBERTa v3 serves as a reward model to assess answer reliability and guide PPO optimization; data Shapley values quantify the contribution of each data point to model performance, weighting it during reward distribution and retraining. In terms of model training, DecideAI employs the RLHF method combined with the PPO algorithm, maximizing the expected performance of the model while ensuring stability; this process not only enhances the model's performance on specific objectives compared to traditional strategies but also significantly reduces the risk of hallucinations and erroneous answers.
Difference from centralized LLM development
Unlike traditional closed large model development dominated by a few giants, DecideAI emphasizes open collaboration and data quality. Through the PoP and Decide ID mechanisms, each training data can be traced back to real and trustworthy contributors; the system uses methods like data Shapley to evaluate and reward high-value data contributions. The native token DCD and built-in reward system ensure contributor motivation, allowing ecological participants to benefit from their contributions. Meanwhile, technologies like zero-knowledge proofs ensure the privacy and security of participants. In contrast, traditional closed systems often pursue only the scale of data and models, neglecting the verifiability of data sources and the customization needs of models. DecideAI focuses on 'quality over quantity', achieving higher customization and performance through transparent collaboration processes, opening up a new paradigm for development.
In summary, DecideAI combines blockchain and AI technology to build a decentralized training and sharing ecology, providing innovative solutions for high-quality data supply and professional model collaboration in the AI era.