Model Extrapolation Expedites Alignment
Fuente:
arXiv
Enregistré dans:
| Auteurs principaux: | , , , , |
|---|---|
| Format: | Preprint |
| Publié: |
2024
|
| Sujets: | |
| Accès en ligne: | |
| Tags: |
Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
|
| _version_ | 1866908384441139200 |
|---|---|
| author | Zheng, Chujie Wang, Ziqi Ji, Heng Huang, Minlie Peng, Nanyun |
| author_facet | Zheng, Chujie Wang, Ziqi Ji, Heng Huang, Minlie Peng, Nanyun |
| contents | Given the high computational cost of preference alignment training of large language models (LLMs), exploring efficient methods to reduce the training overhead remains an important and compelling research problem. Motivated by the observation that alignment training typically involves only small parameter changes without injecting new knowledge into models, we propose a straightforward method called ExPO (model extrapolation) to expedite LLMs' alignment with human preferences. Given a partially-trained model and its initial SFT checkpoint, ExPO improves the implicit optimization objective of alignment training by simply amplifying the parameter change based on a first-order approximation, without any additional training overhead. Through controlled experiments, we demonstrate that ExPO boosts a DPO model trained with only 20% steps to outperform the fully-trained one. Moreover, we show that ExPO notably improves existing open-source LLMs (ranging from 1.8B to 70B parameters) on the leading AlpacaEval 2.0 and MT-Bench benchmarks, which highlights ExPO's broader utility in efficiently enhancing LLM alignment. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2404_16792 |
| institution | arXiv |
| publishDate | 2024 |
| record_format | arxiv |
| spellingShingle | Model Extrapolation Expedites Alignment Zheng, Chujie Wang, Ziqi Ji, Heng Huang, Minlie Peng, Nanyun Machine Learning Artificial Intelligence Computation and Language Given the high computational cost of preference alignment training of large language models (LLMs), exploring efficient methods to reduce the training overhead remains an important and compelling research problem. Motivated by the observation that alignment training typically involves only small parameter changes without injecting new knowledge into models, we propose a straightforward method called ExPO (model extrapolation) to expedite LLMs' alignment with human preferences. Given a partially-trained model and its initial SFT checkpoint, ExPO improves the implicit optimization objective of alignment training by simply amplifying the parameter change based on a first-order approximation, without any additional training overhead. Through controlled experiments, we demonstrate that ExPO boosts a DPO model trained with only 20% steps to outperform the fully-trained one. Moreover, we show that ExPO notably improves existing open-source LLMs (ranging from 1.8B to 70B parameters) on the leading AlpacaEval 2.0 and MT-Bench benchmarks, which highlights ExPO's broader utility in efficiently enhancing LLM alignment. |
| title | Model Extrapolation Expedites Alignment |
| topic | Machine Learning Artificial Intelligence Computation and Language |
| url | https://arxiv.org/abs/2404.16792 |