Summary of Does Your Data Spark Joy? Performance Gains From Domain Upsampling at the End Of Training, by Cody Blakeney et al.
Does your data spark joy? Performance gains from domain upsampling at the end of training
by Cody Blakeney, Mansheej Paul, Brett W. Larsen, Sean Owen, Jonathan Frankle
First submitted to arxiv on: 5 Jun 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Computation and Language (cs.CL)
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 a technique to optimize the balance between general web scrape data and domain-specific data for large language models (LLMs). It shows that by upsampling smaller domain-specific datasets relative to CommonCrawl (CC) at the end of training, performance improvements can be achieved on difficult benchmarks. The method rivals the results of Llama-2, a 7B model trained for twice as long. The paper also experiments with ablating the duration of domain upsampling and finds that 10% to 20% is optimal. Additionally, it demonstrates how to remove individual datasets during the final phase of training to characterize their utility at scale. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary The paper helps us understand how to use smaller domain-specific datasets to improve the capabilities of large language models (LLMs). It does this by upsampling these datasets relative to general web scrape data. This technique can be used to get better results on difficult tasks, and it’s a more affordable way to do so than training the model from scratch. |
Keywords
» Artificial intelligence » Llama
