Smaug: Fixing Failure Modes of Preference Optimisation with DPO-Positive
- URL: http://arxiv.org/abs/2402.13228v2
- Date: Wed, 3 Jul 2024 13:46:33 GMT
- Title: Smaug: Fixing Failure Modes of Preference Optimisation with DPO-Positive
- Authors: Arka Pal, Deep Karkhanis, Samuel Dooley, Manley Roberts, Siddartha Naidu, Colin White,
- Abstract summary: We show theoretically that the standard DPO loss can lead to a reduction of the model's likelihood of the preferred examples.
We design DPO-Positive (DPOP), a new loss function and training procedure which avoids this failure mode.
Surprisingly, we find that DPOP outperforms DPO and other fine-tuning procedures across a wide variety of datasets and downstream tasks.
- Score: 15.066029556877721
- License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
- Abstract: Direct Preference Optimisation (DPO) is effective at significantly improving the performance of large language models (LLMs) on downstream tasks such as reasoning, summarisation, and alignment. Using pairs of preferred and dispreferred data, DPO models the relative probability of picking one response over another. In this work, first we show theoretically that the standard DPO loss can lead to a reduction of the model's likelihood of the preferred examples, as long as the relative probability between the preferred and dispreferred classes increases. We then show empirically that this phenomenon occurs when fine-tuning LLMs on common datasets, especially datasets in which the edit distance between pairs of completions is low. Using these insights, we design DPO-Positive (DPOP), a new loss function and training procedure which avoids this failure mode. Surprisingly, we find that DPOP outperforms DPO and other fine-tuning procedures across a wide variety of datasets and downstream tasks, including datasets with high edit distances between completions. Furthermore, we find that the DPOP-tuned model outperforms the DPO-tuned model (all else equal) on benchmarks independent of the fine-tuning data, such as MT-Bench. Finally, using DPOP, we create and open-source Smaug-34B and Smaug-72B, with the latter becoming the first open-source LLM to surpass an average accuracy of 80% on the HuggingFace Open LLM Leaderboard.
Related papers
- Step-DPO: Step-wise Preference Optimization for Long-chain Reasoning of LLMs [54.05511925104712]
We propose a simple, effective, and data-efficient method called Step-DPO.
Step-DPO treats individual reasoning steps as units for preference optimization rather than evaluating answers holistically.
Our findings demonstrate that as few as 10K preference data pairs and fewer than 500 Step-DPO training steps can yield a nearly 3% gain in accuracy on MATH for models with over 70B parameters.
arXiv Detail & Related papers (2024-06-26T17:43:06Z) - On Softmax Direct Preference Optimization for Recommendation [50.896117978746]
We propose Softmax-DPO to instill ranking information into the LM to help LM-based recommenders distinguish preferred items from negatives.
Specifically, we incorporate multiple negatives in user preference data and devise an alternative version of DPO loss tailored for LM-based recommenders.
arXiv Detail & Related papers (2024-06-13T15:16:11Z) - Self-Augmented Preference Optimization: Off-Policy Paradigms for Language Model Alignment [104.18002641195442]
We introduce Self-Augmented Preference Optimization (SAPO), an effective and scalable training paradigm that does not require existing paired data.
Building on the self-play concept, which autonomously generates negative responses, we further incorporate an off-policy learning pipeline to enhance data exploration and exploitation.
arXiv Detail & Related papers (2024-05-31T14:21:04Z) - Robust Preference Optimization through Reward Model Distillation [68.65844394615702]
Language model (LM) post-training involves maximizing a reward function that is derived from preference annotations.
DPO is a popular offline alignment method that trains a policy directly on preference data without the need to train a reward model or apply reinforcement learning.
We analyze this phenomenon and propose distillation to get a better proxy for the true preference distribution over generation pairs.
arXiv Detail & Related papers (2024-05-29T17:39:48Z) - Triple Preference Optimization: Achieving Better Alignment with Less Data in a Single Step Optimization [35.36615140853107]
Triple Preference Optimization (TPO) is designed to align large language models with three preferences without requiring a separate Supervised Fine-Tuned (SFT) model.
We show that TPO achieves superior results compared to models aligned through other methods such as SFT, DPO, KTO, IPO, CPO, and ORPO.
arXiv Detail & Related papers (2024-05-26T20:18:11Z) - Multi-Reference Preference Optimization for Large Language Models [56.84730239046117]
We introduce a novel closed-form formulation for direct preference optimization using multiple reference models.
The resulting algorithm, Multi-Reference Preference Optimization (MRPO), leverages broader prior knowledge from diverse reference models.
Our experiments demonstrate that LLMs finetuned with MRPO generalize better in various preference data, regardless of data scarcity or abundance.
arXiv Detail & Related papers (2024-05-26T00:29:04Z) - Self-Play Preference Optimization for Language Model Alignment [75.83359213697854]
Recent advancements suggest that directly working with preference probabilities can yield a more accurate reflection of human preferences.
We propose a self-play-based method for language model alignment, which treats the problem as a constant-sum two-player game.
Our approach, dubbed Self-Play Preference Optimization (SPPO), approximates the Nash equilibrium through iterative policy updates.
arXiv Detail & Related papers (2024-05-01T17:59:20Z) - Mixed Preference Optimization: Reinforcement Learning with Data Selection and Better Reference Model [3.300814846990438]
Large Language Models (LLMs) have become increasingly popular due to their ability to process and generate natural language.
As they are trained on massive datasets of text, LLMs can inherit harmful biases and produce outputs that are not aligned with human values.
This paper studies two main approaches to LLM alignment: Reinforcement Learning with Human Feedback (RLHF) and contrastive learning-based methods like Direct Preference Optimization (DPO)
By analyzing the stability and robustness of RLHF and DPO, we propose MPO, a novel method that mitigates the weaknesses of both approaches.
arXiv Detail & Related papers (2024-03-28T14:15:10Z) - sDPO: Don't Use Your Data All at Once [11.149898528381902]
We propose stepwise DPO (sDPO), an extension of the recently popularized direct preference optimization (DPO) for alignment tuning.
This approach involves dividing the available preference datasets and utilizing them in a stepwise manner, rather than employing them all at once.
We demonstrate that this method facilitates the use of more precisely aligned reference models within the DPO training framework.
arXiv Detail & Related papers (2024-03-28T09:56:04Z) - Statistical Rejection Sampling Improves Preference Optimization [42.57245965632205]
We introduce a novel approach to source preference data from the target optimal policy using rejection sampling.
We also propose a unified framework that enhances the loss functions used in both Sequence Likelihood (SLiC) and Direct Preference Optimization (DPO) from a preference modeling standpoint.
arXiv Detail & Related papers (2023-09-13T01:07:25Z)
This list is automatically generated from the titles and abstracts of the papers in this site.
This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.