Constraint Back-translation Improves Complex Instruction Following of Large Language Models
Yunjia Qi, Hao Peng, Xiaozhi Wang, Bin Xu, Lei Hou, Juanzi Li
2024-11-01
Summary
This paper discusses a new method called constraint back-translation that helps large language models (LLMs) better follow complex instructions by improving the quality of training data without needing a lot of human input.
What's the problem?
Large language models often struggle to follow complex instructions that have specific requirements, like format or length. Previous methods for training these models relied on creating instruction-response pairs, but even advanced models can have trouble with complex instructions, leading to poor-quality data and limited performance.
What's the solution?
The authors propose a technique called constraint back-translation, which uses existing high-quality instruction-response pairs from datasets. Instead of generating new complex instructions, they identify the implicit constraints already met by the responses and add these constraints to the original instructions. This method reduces the need for extensive human annotations and improves data quality. They create a new dataset called CRAB using this approach and show that training on this dataset enhances the ability of various LLMs to follow complex instructions.
Why it matters?
This research is important because it provides a more efficient way to improve LLMs' performance on complex tasks without requiring a lot of additional resources. By enhancing how these models understand and follow intricate instructions, it can lead to better applications in areas like customer service, education, and any field where precise communication is essential.
Abstract
Large language models (LLMs) struggle to follow instructions with complex constraints in format, length, etc. Following the conventional instruction-tuning practice, previous works conduct post-training on complex instruction-response pairs generated by feeding complex instructions to advanced LLMs. However, even advanced LLMs cannot follow complex instructions well, thus limiting the quality of generated data. In this work, we find that existing datasets inherently contain implicit complex constraints and propose a novel data generation technique, constraint back-translation. Specifically, we take the high-quality instruction-response pairs in existing datasets and only adopt advanced LLMs to add complex constraints already met by the responses to the instructions, which naturally reduces costs and data noise. In the experiments, we adopt Llama3-70B-Instruct to back-translate constraints and create a high-quality complex instruction-response dataset, named CRAB. We present that post-training on CRAB improves multiple backbone LLMs' complex instruction-following ability, evaluated on extensive instruction-following benchmarks. We further find that constraint back-translation also serves as a useful auxiliary training objective in post-training. Our code, data, and models will be released to facilitate future research.