English

ORPO: Monolithic Preference Optimization without Reference Model

Computation and Language 2024-03-15 v2 Artificial Intelligence

Abstract

While recent preference alignment algorithms for language models have demonstrated promising results, supervised fine-tuning (SFT) remains imperative for achieving successful convergence. In this paper, we study the crucial role of SFT within the context of preference alignment, emphasizing that a minor penalty for the disfavored generation style is sufficient for preference-aligned SFT. Building on this foundation, we introduce a straightforward and innovative reference model-free monolithic odds ratio preference optimization algorithm, ORPO, eliminating the necessity for an additional preference alignment phase. We demonstrate, both empirically and theoretically, that the odds ratio is a sensible choice for contrasting favored and disfavored styles during SFT across the diverse sizes from 125M to 7B. Specifically, fine-tuning Phi-2 (2.7B), Llama-2 (7B), and Mistral (7B) with ORPO on the UltraFeedback alone surpasses the performance of state-of-the-art language models with more than 7B and 13B parameters: achieving up to 12.20% on AlpacaEval2.0\text{AlpacaEval}_{2.0} (Figure 1), 66.19% on IFEval (instruction-level loose, Table 6), and 7.32 in MT-Bench (Figure 12). We release code and model checkpoints for Mistral-ORPO-α\alpha (7B) and Mistral-ORPO-β\beta (7B).

Keywords

Cite

@article{arxiv.2403.07691,
  title  = {ORPO: Monolithic Preference Optimization without Reference Model},
  author = {Jiwoo Hong and Noah Lee and James Thorne},
  journal= {arXiv preprint arXiv:2403.07691},
  year   = {2024}
}

Comments

Preprint

R2 v1 2026-06-28T15:17:21.486Z