Summary of Diversify and Conquer: Diversity-centric Data Selection with Iterative Refinement, by Simon Yu and Liangyu Chen and Sara Ahmadian and Marzieh Fadaee
Diversify and Conquer: Diversity-Centric Data Selection with Iterative Refinement
by Simon Yu, Liangyu Chen, Sara Ahmadian, Marzieh Fadaee
First submitted to arxiv on: 17 Sep 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 This paper proposes an innovative approach to finetuning large language models (LLMs) on instruction data, aiming to enhance pre-trained knowledge and improve instruction-following capabilities. By recognizing the importance of selecting optimal data for effective training, the authors introduce a global method focused on data diversity, contrasting with existing research that emphasizes local criteria like instance quality. The proposed method employs k-means clustering to ensure the selected subset effectively represents the full dataset and iteratively refines sampling weights based on each cluster’s importance. This approach reduces the effect of outliers and automatically filters out low-quality data. Through extensive evaluation across various tasks, including natural language reasoning, general world knowledge, code and math reasoning, the authors observe consistent improvements, achieving a 7% increase over random selection and a 3.8% improvement over state-of-the-art sampling methods. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper helps us understand how to train language models better. Right now, we’re fine-tuning these models on lots of different instructions, which is important for them to learn new things. The question is: How do we choose the right instruction data so our model gets better? Most people focus on making sure each piece of data is good quality, but this paper says that’s not as important as having a mix of different kinds of data. They came up with a new way to do this using something called k-means clustering. This helps us choose the right data and ignore the bad stuff. When they tested it on many tasks, like understanding language and doing math, it worked really well and made their model better. |
Keywords
» Artificial intelligence » Clustering » Fine tuning » K means
