Overview of the DecideAI project and industry opportunities
In recent years, the large language model (LLM) market has experienced explosive growth, expected to reach $260 billion by 2030. Enterprise-level applications (such as healthcare, finance, e-commerce, and media) are the main driving force, with LLMs optimizing processes and enhancing customer experiences. Meanwhile, the demand for data and infrastructure required to train these models has surged. However, most general large models are currently controlled by a few centralized giants, leading to issues of non-transparency in the training process, potential biases, and privacy risks. The demand for high-quality, customizable models in specific domains is gradually increasing. In a context where the 'low-hanging fruit' is gradually depleting, developing customized models for professional scenarios has become a long-term growth point. DecideAI was born against this backdrop, with the vision of creating a transparent, trustworthy, scalable, and cost-effective open AI training ecosystem for users, contributors, and developers through blockchain and community collaboration technologies. DecideAI believes that traditional closed-source models and generic datasets are constraints to innovation, necessitating a new model that integrates artificial intelligence with blockchain verification to address quality and trust issues.
Structure and core modules of the DecideAI ecosystem
The DecideAI ecosystem consists of three complementary core modules:
Decide Protocol: A training platform focused on 'RLHF (Reinforcement Learning with Human Feedback)', collaborating with artificial intelligence and human experts to annotate, train, and continuously optimize customized LLMs and corresponding datasets.
Decide ID: Utilizing the Proof of Personhood (PoP) method to authenticate the identity and qualifications of all participants, ensuring that each contributor is a real person with the corresponding professional background, thereby guaranteeing the quality of training data.
Decide Cortex: A knowledge-sharing platform that provides developers and enterprises access to pre-trained models and high-quality datasets (available through purchasing licenses or API calls), facilitating users to quickly build or customize models without starting from scratch.
These three components work in synergy to build an open and sustainable AI ecosystem. The architecture of DecideAI incorporates a comprehensive incentive mechanism that encourages long-term participation and deters malicious behavior. In short, DecideAI establishes a new industry standard for AI training data and model development through a quality-first collaborative model and blockchain-based verification mechanism.
Details of the RLHF process and incentive mechanisms of Decide Protocol
Decide Protocol is centered around 'RLHF (Reinforcement Learning with Human Feedback)' as its core methodology, balancing cost-effectiveness and data quality. Its training process typically includes the following five stages:
Seeding Stage: Engaging in dialogue with the foundational model using initial prompts relevant to the target domain, allowing the model to gradually 'familiarize' itself with the domain concepts and form an initial foundational model.
Annotation Stage: Experts with domain qualifications are selected via Decide ID to evaluate and annotate the dialogues generated by the model, providing feedback, supplementary information, and correction suggestions to enhance the accuracy and relevance of the dataset and model responses. During the annotation process, contributors may categorize, score, and rank model outputs to determine the best responses.
Incentivization Stage: For each round of annotation and model improvement, the system rewards contributors using the native token DCD. Token rewards are allocated based on contribution value, with a public leaderboard incentivization mechanism that grants additional rewards to top performers. This mechanism aims to encourage high-quality contributions, improve participant retention, and deter malicious behavior. Additionally, developers can earn DCD rewards by creating tools and applications, promoting ecological innovation and application implementation.
Model Training: Utilizing annotated data to train the model and further optimize it using advanced RL algorithms. Decide Protocol employs the TRLX framework and PPO algorithm to iteratively adjust the reward model output, enhancing model capabilities while maintaining training stability. Furthermore, techniques such as 'data Shapley value' analysis and cross-validation are introduced to further assess each training data point's contribution to model performance and optimize the allocation and quality control of the annotation process.
Continuous Evolution: After completing the initial training, the trained model is compared with the latest cutting-edge models to decide whether to deploy the current version or retrain a new foundational model. After deployment, Decide Protocol continuously collects performance and feedback from the model in real-world applications, constantly identifying improvement points and feeding these insights back into the annotation process to ensure the model remains aligned with actual needs.
The above processes implement a closed loop of machine-human collaborative optimization of model quality through the RLHF framework. The entire process is transparently recorded on the blockchain, with each data point traceable to specific verified human contributors. The token incentive and leaderboard mechanism not only motivate contributors to provide valuable annotations but also encourage long-term participation and contributions to the ecosystem, creating a virtuous cycle.
PoP identity verification and data quality assurance mechanisms of Decide ID
In the Decide Protocol, high-quality training data relies on reliable human resources. The Decide ID module ensures that each participant is a unique, real expert through the 'Proof of Personhood (PoP)' method. The specific process is as follows:
Identity Generation and Authentication: Each new user generates an on-chain identity (Principal ID) upon first login to the platform and then verifies their uniqueness by submitting biometric features (such as fingerprints or facial information). Subsequently, users must also provide additional qualification proofs (such as educational background, work certifications, etc.) to demonstrate their professional competence. Decide ID employs zero-knowledge proof technology to complete identity and qualification verification without exposing sensitive user information, ensuring users receive verification while protecting their privacy.
Application of Verification Results: Only verified users can participate in the data creation and annotation of the Decide Protocol. Thus, each data contribution used for training can be traced back to verified real experts. The system can prevent interference from bots and fake identities through on-chain stored verification records, greatly enhancing the credibility and quality of the training data.
In summary, Decide ID injects a trustworthy 'human factor' into the ecosystem. It integrates identity verification results with the RLHF process through multi-step authentication and blockchain records, creating a closed-loop data quality assurance system.
Decide Cortex's data and model sharing platform
Decide Cortex is positioned as a decentralized model and data-sharing platform, aiming to make the most advanced AI models accessible and promote community collaboration. Its main features include:
Knowledge Sharing: Decide Cortex offers a range of pre-trained models and high-quality datasets, which users can access through purchasing licenses or API calls. The platform supports general large language models (serving as the foundation for various NLP tasks) as well as customized models tailored to specific industries or scenarios, making it convenient for developers and enterprises to select and customize according to their needs.
Customizability and Management: Users can not only use pre-trained models on the platform but also fine-tune and retrain them. The platform provides monitoring and management features, supporting performance monitoring and iterative updates of models after they go live.
Open Collaboration: Decide Cortex encourages users and developers to open-source and share their models and datasets, building a community-driven ecosystem. For teams that do not want to build models from scratch, Decide Cortex enables them to quickly access trustworthy models and data, lowering the barriers to innovation.
Through these functionalities, Decide Cortex returns high-quality models and datasets trained through the Decide Protocol back to the community, achieving the recycling of ecological resources and value sharing.
Summary and Prospects
Overall, the DecideAI project constructs an open, transparent, and quality-oriented AI training ecosystem through the organic combination of blockchain verification for identity and data, RLHF-based model training, and token incentive mechanisms. This system has significant advantages in improving model accuracy, ensuring data credibility, and encouraging community collaboration, representing a cutting-edge exploration of the deep integration of #AI and #Web3 .
According to the official white paper, with the growing demand for high-quality data and specialized models, DecideAI aims to be at the forefront of this trend, leading the development of the next generation of AI ecosystems through a quality-first open collaboration model. Its technology path based on decentralized trust and open innovation provides new ideas and paradigms for the integration of blockchain and artificial intelligence.