Core Insights - The company Xiaohongshu has launched its first social large model "RedOne," which is a customized LLM aimed at enhancing performance in the SNS (Social Network Service) sector [1] - RedOne shows an average improvement of 14.02% across eight major SNS tasks compared to baseline models, and a 7.56% enhancement in bilingual evaluation benchmarks [1] - In online testing, RedOne reduces harmful content detection exposure by 11.23% and increases post-view search click-through rates by 14.95% compared to single-task fine-tuned baseline models [1]

Core Viewpoint - Xiaohongshu has launched RedOne, a customized large language model (LLM) designed for the social networking service (SNS) sector, aiming to enhance performance across various SNS tasks and improve user interaction and content governance [1][5]. Summary by Sections Introduction of RedOne - RedOne is the first social model that integrates social understanding with platform rules, achieving an average performance improvement of 14.02% across eight major SNS tasks compared to baseline models [1][5]. - In online tests, RedOne reduced harmful content exposure by 11.23% and increased post-view search click-through rates by 14.95% [1][32]. Challenges in SNS Data - SNS data is characterized by high non-standardization, strong contextual dependence, and significant emotional variability, posing challenges for traditional NLP platforms [3][4]. - Existing models often focus on single-task performance, leading to diminishing returns and inadequate adaptability in multi-language and multi-cultural contexts [4]. Training Framework of RedOne - RedOne employs a three-stage training strategy: Continue Pretraining (CPT), Supervised Fine-Tuning (SFT), and Preference Optimization (PO) [5][19]. - The model is trained on a large-scale dataset exceeding 100 billion tokens, combining general high-quality data and SNS-specific data [9][12]. Data Collection and Quality Control - Data is collected from high-quality open-source corpora and SNS platforms, ensuring a diverse representation of social communication styles [8][10]. - A rigorous data filtering process is implemented to maintain high quality, resulting in a final dataset of 20 billion tokens for training [11][12]. Supervised Fine-Tuning (SFT) - SFT focuses on aligning the model's capabilities with real-world application needs, incorporating various tasks such as content understanding and user behavior modeling [15][16]. - The training process emphasizes retaining typical SNS language styles to ensure the model's relevance in real-world scenarios [16][17]. Preference Optimization (PO) - PO enhances the model's alignment with human preferences and platform rules by utilizing a preference dataset constructed through expert annotations and structural information from data labels [20][21]. - Direct Preference Optimization (DPO) is employed to refine the model's outputs, improving its adaptability to SNS environments [22]. Performance Comparison - RedOne outperforms baseline models in various tasks, demonstrating significant improvements in both general and SNS-specific benchmarks [23][26]. - The model's performance continues to improve with increased scale, indicating its potential for further advancements [26][28]. Online Experiment Results - In real-world applications, RedOne significantly reduced harmful content exposure and improved user engagement metrics, showcasing its practical value [32]. Generalization and Robustness - Experiments indicate that incorporating general domain data enhances the model's generalization capabilities, particularly in out-of-distribution tasks [35]. Future Outlook - RedOne represents a significant advancement in addressing the challenges of content governance and interaction quality in SNS, providing a reference for future specialized LLM applications [37].