Condor: Enhance LLM Alignment with Knowledge-Driven Data Synthesis and Refinement
This addresses the data scarcity problem for researchers and practitioners fine-tuning advanced LLMs, offering a scalable synthetic data solution that is incremental in improving existing methods.
The authors tackled the bottleneck of high-quality human-annotated data for fine-tuning large language models by introducing Condor, a two-stage synthetic data generation framework that uses world knowledge and self-reflection; their results show that fine-tuning on only 20K Condor-generated samples achieves superior performance compared to counterparts, with iterative self-improvement validated up to 72B scales.
The quality of Supervised Fine-Tuning (SFT) data plays a critical role in enhancing the conversational capabilities of Large Language Models (LLMs). However, as LLMs become more advanced, the availability of high-quality human-annotated SFT data has become a significant bottleneck, necessitating a greater reliance on synthetic training data. In this work, we introduce Condor, a novel two-stage synthetic data generation framework that incorporates World Knowledge Tree and Self-Reflection Refinement to produce high-quality SFT data at scale. Our experimental results demonstrate that a base model fine-tuned on only 20K Condor-generated samples achieves superior performance compared to counterparts. The additional refinement stage in Condor further enables iterative self-improvement for LLMs at various scales (up to 72B), validating the effectiveness of our approach. Furthermore, our investigation into the scaling for synthetic data in post-training reveals substantial unexplored potential for performance improvements, opening promising avenues for future research.