This repository contains technical documents detailing the research and development efforts by OLM Research. Our work focuses on fine-tuning AI models and integrating AI with decentralized platforms to advance the capabilities of the OLM ecosystem.

Researches

R&D of OLM Instruction Fine-Tuning Models

This research outlines the methodologies applied in fine-tuning OpenLM models, including supervised instruction-tuning and reinforcement learning from human feedback (RLHF). It describes the use of diverse datasets, combining human-annotated and GPT-generated data, to enhance model performance for various applications such as conversational AI and evaluation tasks. Fine-tuning techniques include the use of cosine learning rate schedules, teacher-forcing with loss masking, and advanced algorithms like Direct Policy Optimization (DPO) and Rejection Sampling for iterative training.

R&D of SearchOLM in ChatOLM

This document covers the development of SearchOLM, a decentralized search interface integrated with ChatOLM. It provides a technical overview of the system architecture, including search query processing, content retrieval from the web, natural language understanding via ORA's on-chain AI Oracle, and ranking of extracted information. The paper details components such as query rewriting, vector-based content embedding, and response generation, aimed at enabling seamless and decentralized information retrieval within the ChatOLM ecosystem.