Summary of Disperse-then-merge: Pushing the Limits Of Instruction Tuning Via Alignment Tax Reduction, by Tingchen Fu et al.
Disperse-Then-Merge: Pushing the Limits of Instruction Tuning via Alignment Tax Reduction
by Tingchen Fu, Deng Cai, Lemao Liu, Shuming Shi, Rui Yan
First submitted to arxiv on: 22 May 2024
Categories
- Main: Computation and Language (cs.CL)
- Secondary: Artificial Intelligence (cs.AI)
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 proposed disperse-then-merge framework for supervised fine-tuning (SFT) on instruction-following corpus improves the alignment of large language models (LLMs) on standard knowledge and reasoning benchmarks, outperforming sophisticated methods. The framework disperses instruction-following data into portions, trains multiple sub-models using different data portions, and then merges them via model merging techniques to address the phenomenon of alignment tax. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Large language models can be aligned with human knowledge by fine-tuning them on instruction-following data. However, this process often results in a decline in performance on standard benchmarks. Researchers found that data biases might be the cause of this issue and developed a simple solution. They split the data into smaller parts, trained separate models on each part, and then combined the models. This approach worked better than more complex methods. |
Keywords
» Artificial intelligence » Alignment » Fine tuning » Supervised
