LIONs: An Empirically Optimized Approach to Align Language Models
This work addresses the lack of comprehensive studies on alignment design choices for language models, offering incremental improvements through empirical optimization.
The authors tackled the problem of aligning language models for instruction-following by empirically optimizing a three-stage training pipeline, finding that techniques like sequence packing and increasing dataset size in DPO led to models that outperformed official instruct models from Gemma-2b-base and LLama-3-8b-base.
Alignment is a crucial step to enhance the instruction-following and conversational abilities of language models. Despite many recent work proposing new algorithms, datasets, and training pipelines, there is a lack of comprehensive studies measuring the impact of various design choices throughout the whole training process. We first conduct a rigorous analysis over a three-stage training pipeline consisting of supervised fine-tuning, offline preference learning, and online preference learning. We have found that using techniques like sequence packing, loss masking in SFT, increasing the preference dataset size in DPO, and online DPO training can significantly improve the performance of language models. We then train from Gemma-2b-base and LLama-3-8b-base, and find that our best models exceed the performance of the official instruct models tuned with closed-source data and algorithms. Our code and models can be found at \url{https://github.com/Columbia-NLP-Lab/LionAlignment}.