RedOne - A social media platform launched by Xiaohongshu

What is RedOne?

RedOne is Xiaohongshu's first customized Large Language Model (LLM) for social networking services (SNS). The model employs a three-stage training strategy, incorporating social culture knowledge to enhance multi-task capabilities and aligning with platform norms and human preferences. Compared to the base model, RedOne achieves an average performance improvement of 14.02% across eight social tasks and a 7.56% improvement on bilingual benchmarks. The model reduces exposure by 11.23% in harmful content detection and increases page click-through rate by 14.95% in post-browse search. RedOne demonstrates outstanding performance in the social domain, providing robust support for SNS applications.

RedOne's main functions

• Content ComprehensionIt can classify, identify themes, and understand intent in user-generated content.
• Information ExtractionExtract structured information from informal social posts, such as predicting tags, answering questions, and detecting key words.
• Semantic matchingDetermine the semantic relationship between user queries and social notes, and provide a relevance assessment.
• User behavior modelingSimulate user behavior, such as generating subsequent queries based on browsing history.
• Dialogue and Role SimulationSupports role-playing in emotional companionship conversations and group chats.
• translateTranslate notes in a multilingual environment while preserving the original tone and emotion.
• Harmful content detectionReduce the exposure of harmful content and improve platform security.
• Search optimization after browsingIncrease user click-through rate and enhance content discovery capabilities.

RedOne's technical principles

• Continue pretraining (CPT)RedOne's continued pre-training phase infuses the model with fundamental knowledge of the social domain. Researchers collected large-scale data from general high-quality corpora and social network platforms, covering various social communication patterns such as informal discussions, short comments, and sarcastic remarks. Using a carefully designed data filtering process, low-quality data was removed and the data mixture distribution was optimized, allowing the model to be trained further based on Qwen2.5.
• Supervised Fine-Tuning (SFT)In the supervised fine-tuning phase, based on carefully designed task definitions and data construction, the gap between pre-training objectives and actual SNS application needs is bridged. Researchers compiled a large amount of real user-generated content, defined six core capabilities, including content understanding, information extraction, and semantic matching, and mapped each capability to specific tasks. Based on a two-step training strategy, the first step involves training with a mixture of large-scale general data and SNS data, and the second step increases the proportion of SNS data to further optimize the model's performance on key tasks.
• Preference Optimization (PO)The preference optimization stage utilizes implicit preference signals to make the model output more consistent with human preferences and platform specifications. Researchers constructed different preference pairs using different strategies based on different task types (subjective and objective tasks), invited experts to label preferences, and expanded the dataset. Based on the Direct Preference Optimization (DPO) algorithm, the model's output is optimized using signals from the preference dataset to make it closer to human preferences.
• Data Hybridization and Preservation of General CapabilitiesDuring training, general domain data and SNS domain data are mixed to maintain the model's generality while improving its adaptability in the SNS domain, enhancing the model's performance in specific domains, and improving its generalization ability on unseen tasks (Out-of-Domain, OOD).

RedOne's project address

• arXiv technical paper: https://www.arxiv.org/pdf/2507.10605

RedOne application scenarios

• Harmful content detectionEffectively identify and filter harmful content, such as hate speech, misinformation, pornography, and violence. By reducing the exposure of harmful content, significantly improve platform security and user experience.
• Search optimization after browsingBased on users' browsing history and behavior, more accurate search suggestions and recommended content can be generated, which can improve users' content discovery capabilities and enhance user interaction with the platform.
• Content comprehension and classificationIt automatically categorizes and understands user-generated content, helping the platform to better manage and recommend content.
• Information ExtractionExtracting key information, such as tags, keywords, and key facts, from informal social media posts is extremely helpful for content recommendation, information aggregation, and knowledge graph construction.
• Semantic matching: Evaluate the semantic relevance between user queries and social notes to provide more accurate search results and recommended content.