OpenLedger 백서

OpenLedger

원문

bitpro가 원문에서 글자만 모은 내용이에요. 표·그림·수식은 빠지거나 깨질 수 있어요. 전체 2쪽 중 2쪽

purposeful and specialized AI agents . 3 Ecosystem Tools 3.1 Model Factory ModelFactory is an advanced fine-tuning platform for Large Language Mod- els (LLMs) within the OpenLedger ecosystem. Unlike traditional fine-tuning approaches that require extensive command-line expertise or API integra- tions, ModelFactory offers a purely GUI-based solution that integrates se- cure dataset access and automated fine-tuning workflows. By leveraging OpenLedger’s decentralized dataset repository, it ensures that fine-tuning processes are verifiable and secure. 3.1.1 System Architecture The architecture of ModelFactory is modular and designed for scalability. It consists of the following key components: • User Management Module: Manages authentication and dataset access permissions. • Dataset Access Control: Enforces permissioned access to Open- Ledger’s dataset repository. • Fine-Tuning Engine: Supports advanced fine-tuning optimizations such as LoRA and QLoRA. • Chat Interface Module: Enables real-time interaction with fine- tuned models. 14 • RAG Attribution Module: Integrates retrieval-augmented genera- tion (RAG) techniques for source attribution. • Evaluation & Deployment Module: Provides benchmarking, vali- dation, and deployment functionalities. 3.1.2 Fine-Tuning Process ModelFactory employs a structured workflow that begins with dataset re- quest and approval, followed by model selection, training, evaluation, and deployment. Given a dataset D consisting of input-output pairs ( xi, yi), fine-tuning follows an optimization function: θ∗ = arg min θ 1 n nX i=1 L(yi, fθ(xi)) (10) where L is the loss function and θ are the model parameters. The system supports both full-model fine-tuning and parameter-efficient approaches like LoRA, where only adapter weights A and B are updated: y = xW + xAB (11) 3.1.3 Benchmarking and Performance Optimization To evaluate fine-tuned models, ModelFactory provides automated bench- marking based on: • Perplexity and loss reduction on test datasets. • Rouge and BLEU scores for text-based applications. • GPU memory efficiency analysis, particularly for QLoRA models. 3.2 OpenLoRA Open LoRA is a highly efficient framework designed for serving thousands of fine-tuned LoRA (Low-Rank Adaptation) models on a single GPU. It op- timizes resource utilization by enabling dynamic adapter loading, reducing memory overhead, and ensuring high throughput with low latency. Open 15 LoRA is particularly beneficial for applications that require rapid model switching and efficient inference without deploying separate instances for each fine-tuned model. 3.2.1 Key Features • Dynamic Adapter Loading: Just-in-time (JIT) loading of LoRA adapters from OpenLedger • Efficient Memory Utilization: Supports merging adapters per re- quest for ensemble inference without preloading all models into mem- ory. • Optimized Inference: Uses advanced optimizations like tensor paral- lelism, flash-attention, paged attention, and quantization for improved efficiency. • Scalability: Supports serving thousands of fine-tuned LoRA models on a single GPU. • Cost Reduction: Reduces serving costs while maintaining low latency and high throughput. • Streaming & Quantization: Implements token streaming and quan- tization for optimized inference . 4 Model Lifecycle in OpenLedger The OpenLedger model lifecycle is a structured process that ensures AI mod- els are collaboratively developed, economically sustained, and seamlessly in- tegrated into AI-driven applications. Each phase plays a crucial role in build- ing specialized, high-performance AI models while maintaining decentralized governance and financial incentives. 4.1 Model Proposal Developers initiate the process by submitting a model proposal that outlines the model’s purpose, architecture, and intended use case. To ensure com- mitment and prevent spam submissions, proposers may be required to stake a minimum amount of tokens. 16 4.2 Model Governance Model proposals are reviewed and selected through a governance process led by Protocol Governors. Voting power is determined by the number of gOPEN tokens each participant holds. Proposals that reach the required support threshold advance to the next phase of development. This ensures that model progression reflects collective interest and aligns with the incentives of active participants. 4.3 Specialized Data Collection High-quality, domain-specific data is essential for training specialized AI models. OpenLedger facilitates decentralized data collection, ensuring that contributors are rewarded based on the quality and relevance of their data. Data is attributed using cryptographic mechanisms, providing transparency and verifiability while preventing low-quality or adversarial submissions. 4.4 Model Fine-Tuning After data collection, the model undergoes fine-tuning to enhance its perfor- mance for specific applications. This process ensures that the model is more efficient, accurate, and well-suited for real-world deployment. 4.5 Model Optimization and Alignment using RLHF To further refine model behavior, reinforcement learning with human feed- back (RLHF) is applied. Human validators provide feedback on model out- puts, helping to align the model with desired ethical, logical, and functional standards. Contributors who provide valuable feedback are rewarded, while low-quality contributions may face penalties. 4.6 API and Integrations with Agent Frameworks The final stage involves integrating the trained model into AI-driven appli- cations and intelligent agent frameworks. OpenLedger provides APIs that allow developers to connect models and also provide integrations with agent frameworks. These integrations enable AI models to function as decision- making engines in decentralized applications, expanding their usability. 17 5 OPEN Token The OpenLedger Token (OPEN) serves as the fundamental unit of value within the OpenLedger ecosystem. It drives the economic incentives, ensur- ing that all participants, from data contributors to model creators, are fairly rewarded. 5.1 Tokenomics Category Allocation(in percentage) Community 51.71% Investors 18.29% Team 15% Liquidity 5% Ecosystem 10% Table 2: Tokenomics of OPEN Token 5.2 Key Utilities of the Token 5.2.1 Proposal and Platform Fees • Model Creators use tokens to propose new AI models. • A platform fee is paid to the OpenLedger Treasury to ensure long-term sustainability. • The Treasury allocates attribution rewards to contributors. 5.2.2 Data Contribution and Attribution Rewards • Users, Subject Matter Experts (SMEs) and Enterprises contribute data for AI model training. • Contributors earn attribution rewards in tokens based on their data’s impact. 18 5.2.3 Model Creation and Public Hosting • Once enough data is collected and the conditions for the bonding curve are reached, the AI model is created and optimized. • Successfully deployed models generate revenue, which is shared among stakeholders. 5.2.4 Model Inference Payments • Every time an AI model is used for inference, the computation is paid for using tokens. • Fine-tuning occurs via supervised learning and reinforcement learning using human feedback. 6 Making a Self-Sustainable Decentralized AI Ecosystem OpenLedger employs a unified growth flywheel model to drive the sustain- able and scalable development of its AI and blockchain ecosystems. This integrated approach ensures that both ecosystems operate in synergy, rein- forcing each other within a self-perpetuating loop. 6.1 AI Ecosystem Flywheel The growth of the AI ecosystem in OpenLedger begins with model creators and developers who propose, collect specialized data, fine-tune, and deploy specialized AI models. These creators leverage OpenLedger’s Datanets, gaining access to high-quality datasets and secure fine-tuning tools such as ModelFactory and OpenLoRA. As these AI models evolve, they contribute to the creation of a self- sustaining economy, where attribution rewards incentivize further de- velopment and data contributions. Increased model usage ensures that data contributors and model builders are fairly rewarded , reinforcing the ecosystem’s continuous expansion and innovation. 19 6.2 Blockchain Ecosystem Flywheel The blockchain ecosystem in OpenLedger complements and amplifies the AI ecosystem. Built on a decentralized infrastructure, it provides the trust, transparency, and incentive mechanisms required to sustain the platform. 1. Enabling Transparency and Verifiability – OpenLedger trans- forms AI models from black-box systems into transparent, attributable, and verifiable entities, ensuring all contributions and model evolutions can be publicly audited. 2. Governance for Model Advancement – Governance ensure that only the most valuable and community-backed models progress, pre- venting unnecessary development and aligning economic incentives with real-world demand. 3. Model Utilization and Transactions: As models are deployed, users interact with them through the blockchain. These interactions generate transactions recorded immutably on the blockchain, ensuring accountability and data provenance. 4. Increased Validator Engagement: With higher transaction vol- umes and revenue, validators are incentivized to maintain the blockchain infrastructure. This improves network stability and scalability, further attracting developers to build on OpenLedger. 5. Enhanced Ecosystem Trust: The stable and decentralized blockchain fosters trust among participants, encouraging further development, in- novation, and ecosystem expansion. 6.3 Synergy Between AI and Blockchain Ecosystems The synergy between the AI and blockchain ecosystems forms the core of OpenLedger’s growth flywheel. Developers serve as the linchpin, driving innovation in both ecosystems: • From Blockchain to AI: Transparent community enabled model de- velopment process ensure that data and AI models are built on a foun- dation of trust and accountability. The decentralized nature of the blockchain enables a fair and secure platform for AI development. 20 • From AI to Blockchain: As AI models become more sophisticated and widely used, they generate increased revenue and transactions on the blockchain. This strengthens the infrastructure and incentivizes further validator participation, creating a virtuous cycle. • Mutual Reinforcement: Revenue generated from blockchain trans- actions funds further AI innovation, while advancements in AI drive greater adoption and engagement with the blockchain. Together, these ecosystems create a powerful loop of mutual reinforcement. 7 Glossary • Openledger: A decentralized next-gen Blockchain network for AI that provides trust infrastructure for specialized AI models. • OPEN Token: The fundamental token of the OpenLedger ecosystem, used for staking, governance, and transactions. • Specialized Models: AI models fine-tuned for niche tasks using high- quality datasets from OpenLedger’s Datanets. • ModelFactory: A GUI-based platform for secure and efficient fine- tuning of AI models. • OpenLoRA: A multi-tenant system for serving fine-tuned LoRA mod- els with minimal overhead. • Datanets: A secure repository for collecting and managing datasets used for training specialized models. • Retrieval-Augmented Generation (RAG): A technique for gener- ating AI outputs with citations and source attribution. • Proof of Attribution: A system for tracking and verifying the con- tributions of data providers and model creators. • Flywheel Effect: A self-reinforcing cycle of growth where increased participation and innovation drive further adoption and development. • Multi-Tenant GPU Infrastructure: A system that enables efficient sharing of GPU resources across multiple AI models. 21 • Segmented Gather Matrix-Vector Multiplication (SGMV):An optimized CUDA kernel for accelerating LoRA model execution. • Decentralized Governance: A decision-making framework where stakeholders participate in proposals and voting. • Tokenomics: The economic structure of the OPL token, including supply, distribution, and utility. • DataInf: An efficient method for computing data influence in AI mod- els. 8 References References [1] Zhao, B., et al. (2024). ”DataInf: Efficiently Estimating Data In- fluence in LoRA-Tuned LLMs and Diffusion Models.” ICLR 2024. https://arxiv.org/abs/2310.00902 [2] Chen, L., et al. (2023). ”Punica: Multi-Tenant LoRA Serving.” arXiv:2310.18547. https://arxiv.org/abs/2310.18547 22
프로젝트로 돌아가기