RedOne: Revealing Domain-specific LLM Post-Training in Social Networking Services

TL;DR

RedOne addresses the challenge of building a domain-specific LLM for SNS by proposing a three-stage post-training pipeline: continue pretraining (CPT) to acquire SNS-domain knowledge, supervised fine-tuning (SFT) to align with SNS tasks, and preference optimization (PO) via direct preference optimization (DPO) to mirror human preferences. The model is trained on a large-scale dataset combining general and SNS-specific data, filtered with RegMix to produce a high-quality 20B-token SFT set, while CPT uses a 100B-token-scale corpus. Empirical results show RedOne achieves up to 14.02% average improvement on 8 major SNS tasks and 7.56% on SNS-Bench bilingual benchmarks, with online testing revealing an 11.23% reduction in exposure to harmful content and a 14.95% increase in post-view search CTR, all while preserving general capabilities. These findings establish RedOne as a robust, generalizable SNS foundation model with clear practical implications for real-world social platforms and a baseline for future domain-specific LLM research.

Abstract

As a primary medium for modern information dissemination, social networking services (SNS) have experienced rapid growth, which has proposed significant challenges for platform content management and interaction quality improvement. Recently, the development of large language models (LLMs) has offered potential solutions but existing studies focus on isolated tasks, which not only encounter diminishing benefit from the data scaling within individual scenarios but also fail to flexibly adapt to diverse real-world context. To address these challenges, we introduce RedOne, a domain-specific LLM designed to break the performance bottleneck of single-task baselines and establish a comprehensive foundation for the SNS. RedOne was developed through a three-stage training strategy consisting of continue pretraining, supervised fine-tuning, and preference optimization, using a large-scale real-world dataset. Through extensive experiments, RedOne maintains strong general capabilities, and achieves an average improvement up to 14.02% across 8 major SNS tasks and 7.56% in SNS bilingual evaluation benchmark, compared with base models. Furthermore, through online testing, RedOne reduced the exposure rate in harmful content detection by 11.23% and improved the click page rate in post-view search by 14.95% compared with single-tasks finetuned baseline models. These results establish RedOne as a robust domain-specific LLM for SNS, demonstrating excellent generalization across various tasks and promising applicability in real-world scenarios.

Figures (6)

• Figure 1: Performance comparison of different models in the SNS domain, where all models are instruction-tuned and the evaluation score is the average of all tasks on SNS-Bench.
• Figure 2: Overview of our training pipeline.
• Figure 3: Model capability radar diagram across different task categories.
• Figure 4: Performance on OOD tasks for models of varying parameter size.
• Figure 5: Token length distribution in the dataset. The histogram uses a logarithmic y-axis with dashed lines indicating the median (345 tokens) and the 95-th percentile (2,342 tokens).