Summary of Autoscale: Automatic Prediction Of Compute-optimal Data Composition For Training Llms, by Feiyang Kang et al.
AutoScale: Automatic Prediction of Compute-optimal Data Composition for Training LLMs
by Feiyang Kang, Yifan Sun, Bingbing Wen, Si Chen, Dawn Song, Rafid Mahmood, Ruoxi Jia
First submitted to arxiv on: 29 Jul 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Artificial Intelligence (cs.AI); Computation and Language (cs.CL); Machine Learning (stat.ML)
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 explores domain reweighting, a technique that adjusts the relative weights of different data sources to improve language model pre-training. The optimal composition of these domains is shown to be scale-dependent, challenging current practices that rely on small-scale experiments and applying them at larger scales. The authors develop an analytical model for this dependence and introduce AutoScale, a practical approach for optimizing data compositions at large training data scales. This involves using a principled optimization framework to find optimal compositions at smaller scales, then predicting optimal compositions at larger scales using the derived model. The authors evaluate AutoScale on GPT-2 Large and BERT pre-training, demonstrating its effectiveness in improving training convergence and downstream performance. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper looks at how we can make language models better by mixing together different kinds of data. They found that the best mix of data depends on how much data you have, which is a surprise! The authors came up with a new way to figure out the best mix of data called AutoScale. It works by first finding the best mix for a small amount of data and then using that to predict the best mix for a larger amount of data. They tested this on some big language models and found it helped them learn faster and do better on tasks. |
Keywords
» Artificial intelligence » Bert » Gpt » Language model » Optimization
