Summary of Uft: Unifying Fine-tuning Of Sft and Rlhf/dpo/una Through a Generalized Implicit Reward Function, by Zhichao Wang et al.
UFT: Unifying Fine-Tuning of SFT and RLHF/DPO/UNA through a Generalized Implicit Reward Function
by Zhichao Wang, Bin Bi, Zixu Zhu, Xiangbo Mao, Jun Wang, Shiyu Wang
First submitted to arxiv on: 28 Oct 2024
Categories
- Main: Computation and Language (cs.CL)
- Secondary: Machine Learning (cs.LG)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary The paper proposes Unified Fine-Tuning (UFT), a novel method that integrates two stages – Soft Fine-Tuning (SFT) and alignment – into a single training stage using the same objective function. This approach aims to reduce catastrophic forgetting, which occurs when the model forgets previously learned knowledge during the alignment process. Experimental results show that UFT outperforms SFT on instruction-tuning data alone and effectively prevents catastrophic forgetting when combining instruction-tuning data with alignment data. The proposed framework achieves significant improvements in tasks such as instruction-following (ifeval) and factuality (truthful-qa). This work establishes an effective pretraining-UFT paradigm for Large Language Models (LLMs). |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper talks about a new way to train language models, called Unified Fine-Tuning. The goal is to make the model learn from lots of text data and then fine-tune it to do specific tasks like following instructions or answering questions accurately. The problem with previous methods was that they forgot what they learned before, which isn’t good. So, this new method combines two stages into one, making sure the model remembers its past knowledge while still learning new things. This approach works better than others and can help us create more accurate language models. |
Keywords
» Artificial intelligence » Alignment » Fine tuning » Instruction tuning » Objective function » Pretraining
