RAFT: Data Refinement and Adaptive Distillation for Domain Fine-Tuning with Alleviated Forgetting

Researchers introduce RAFT, a framework addressing the problem of catastrophic forgetting in domain-specific fine-tuning of language models. By combining data refinement with answer-conditioned distillation, RAFT achieves 23.2% improvement in domain accuracy while recovering 10-18% of general capability losses typically incurred during fine-tuning.

RAFT addresses a fundamental challenge in machine learning: the trade-off between specialization and generalization. When large language models undergo domain-specific fine-tuning, they typically gain expertise in narrow tasks at the expense of broader capabilities—a phenomenon known as catastrophic forgetting. This research tackles two distinct technical problems that drive this degradation: misalignment between domain training targets and a model's natural response patterns, and the failure of traditional supervised fine-tuning to preserve model behavior on self-generated outputs.

The framework operates in two stages, beginning with data refinement that reframes domain targets to match the model's learned response style through self-conditioned rewriting and semantic filtering. This preprocessing step ensures training signals are compatible with the model's original knowledge. The second stage employs answer-conditioned on-policy distillation, where the base model provides guidance specifically on trajectories generated by the student model, creating a trajectory-level constraint absent in standard approaches.

For practitioners and organizations deploying specialized AI systems, RAFT offers substantial practical value. The 23.2% domain improvement directly translates to better performance on target tasks, while the recovery of general capabilities preserves model utility for broader applications. This matters significantly for enterprises deploying foundation models across multiple use cases, where catastrophic forgetting typically necessitates maintaining separate models or accepting degraded performance.

The research demonstrates consistent improvements across three different base models and five domains, suggesting the approach generalizes well. Future adoption likely hinges on implementation complexity and computational overhead during the distillation phase. Organizations balancing specialized and general-purpose model deployment should monitor developments in trajectory-preserving fine-tuning techniques.

• →RAFT improves domain-specific fine-tuning by 23.2% while recovering 10-18% of lost general capabilities through data refinement and distillation.
• →The framework addresses supervision-compatibility gaps by rewriting domain targets to match model-natural response patterns.
• →Answer-conditioned on-policy distillation preserves model behavior on student-generated trajectories rather than just fixed targets.
• →Results generalize across three instruction-tuned backbones and five domains, indicating robust applicability.
• →Top-K temperature distillation and EMA-based loss balancing provide technical mechanisms for stabilizing domain-general performance trade-offs.