DomainPilot: Domain-Level Loss-Guided Two-Stage Data Mixture Optimization for Efficient Language Model Fine-Tuning DomainPilot：基于域级损失引导的两阶段数据混合优化用于高效语言模型微调

TL;DR

• DomainPilot introduces token-level domain loss monitoring to capture per-domain learning dynamics during training without halting the data pipeline.
• It proposes a Scaling Law guided coarse optimization stage that fits domain-specific convergence curves and derives a principled prior for mixture adjustment.
• A subsequent Mixing Law guided fine optimization stage refines the mixture by modeling cross-domain interaction effects through controlled sweep experiments.
• The framework is realized via a patch-based architecture that injects domain-aware loss computation into existing training frameworks with minimal adapter code.
• Validation on Qwen3-1.7B shows improvements of +2% on MMLU-Redux, +1.8% on AIME24, +3.8% on LiveCodeBench v5, and +3.6% on BFCL v3 without increasing total data volume or training cost.

Why It Matters

DomainPilot addresses critical limitations in dynamic data scheduling methods for industrial-scale pretraining and supervised fine-tuning (SFT) by providing an effective, lightweight alternative to expensive data selection or auxiliary model training. This approach enhances the training efficacy of large language models (LLMs) through optimized data mixture composition, which is crucial for improving performance across various benchmarks without additional computational costs.

Technical Details

• Token-Level Domain Loss Monitoring: Captures per-domain learning dynamics during training without interrupting the data pipeline.
• Scaling Law Guided Coarse Optimization Stage: Fits domain-specific convergence curves to derive a principled prior for mixture adjustment.
• Mixing Law Guided Fine Optimization Stage: Refines the mixture by modeling cross-domain interaction effects through controlled sweep experiments.
• Patch-Based Architecture: Injects domain-aware loss computation into existing training frameworks like MindSpeed/Megatron-LM with approximately 30 lines of framework-specific adapter code.
• Validation: Demonstrated on the Qwen3-1.7B model during SFT, showing significant improvements on multiple benchmarks without increasing total data volume or training cost.

Industry Insight

The development of DomainPilot suggests a promising direction for optimizing data mixture strategies in LLM training, potentially reducing the need for resource-intensive data selection processes and auxiliary model training. For AI professionals, this implies that leveraging domain-level signals can lead to more efficient and effective fine-tuning procedures, ultimately enhancing model performance while maintaining manageable computational requirements.

TL;DR

• 提出DomainPilot框架，通过域级损失引导的两阶段数据混合优化提升大语言模型微调效率。
• 引入Token级域损失监控机制，在不中断数据流的情况下捕捉各领域的学习动态。
• 结合Scaling Law和Mixing Law进行粗粒度与细粒度的混合比例调整，实现跨域交互效应建模。
• 基于Patch架构仅需约30行适配代码即可集成至现有训练框架（如MindSpeed/Megatron-LM）。
• 在Qwen3-1.7B模型SFT实验中，多项基准测试性能显著提升，且未增加数据量或训练成本。

为什么值得看

该研究为工业级LLM高效微调提供了一种轻量、可扩展的数据调度方案，解决了传统方法中高计算开销与I/O瓶颈问题，对资源受限环境下的模型优化具有重要实践价值。其模块化设计便于快速部署，同时揭示了域级信号在数据混合中的关键作用，推动了数据驱动型训练策略的发展。

技术解析

• Token-level domain loss monitoring: 在训练过程中实时统计每个token所属领域的损失值，无需暂停数据流水线即可获取各域的学习进度曲线，避免样本重加权中噪声与难度混淆的问题。
• Two-stage optimization pipeline: 第一阶段基于Scaling Law拟合各域收敛趋势，生成初始混合比例先验；第二阶段通过控制变量扫描实验建模域间相互作用，进一步精细调节配比。
• Patch-based integration mechanism: 采用插件式架构向主流训练框架注入域感知损失计算逻辑，仅需少量特定框架适配器代码（~30 lines），降低工程落地门槛。
• Validation on Qwen3-1.7B SFT: 在保持总数据量和训练时长不变的前提下，优化后的数据混合使MMLU-Redux (+2%)、AIME24 (+1.8%)、LiveCodeBench v5 (+3.8%)、BFCL v3 (+3.6%)等指标全面超越原始混合策略。

行业启示

• 数据质量与结构比单纯扩大规模更能决定模型上限，建议企业在SFT阶段优先投入精力构建领域划分清晰、损失可追踪的训练数据集。
• “无感”监控+渐进式优化的方法论可迁移至其他模态或多任务场景，成为下一代自动化数据编排系统的核心组件雏形。
• 小模型（如1.7B）经此法亦获显著增益，表明该技术不仅适用于超大规模预训练，也适合边缘设备上的垂直领域适配，具备广泛商业化潜力。