Summary of Influential Language Data Selection Via Gradient Trajectory Pursuit, by Zhiwei Deng et al.
Influential Language Data Selection via Gradient Trajectory Pursuit
by Zhiwei Deng, Tao Li, Yang Li
First submitted to arxiv on: 22 Oct 2024
Categories
- Main: Artificial Intelligence (cs.AI)
- Secondary: Computation and Language (cs.CL); 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 This paper proposes a new algorithm called Gradient Trajectory Pursuit (GTP) to curate desirable datasets for training large language models. The authors highlight the limitations of existing methods, which are built upon individual sample rankings or inefficient matching processes, leading to suboptimal performance. GTP jointly selects data points under an L0-norm regularized objective, automatically de-duplicating samples and achieving higher efficiency through compressive sampling processes. The proposed algorithm is demonstrated in both in-domain and target-domain selection benchmarks, showing consistent outperformance of top-k selection and competitive algorithms. For example, GTP chooses as low as 0.5% data to achieve full performance on targeted instruction tuning tasks. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Imagine trying to find the best information for training a super powerful language model. Usually, we look at each piece of information individually or try to match them up, but that’s not very efficient. This paper proposes a new way to do it called Gradient Trajectory Pursuit (GTP). GTP helps us find the right pieces of information by looking at how they work together and getting rid of duplicates. It even works really fast! The researchers tested this method on different tasks and showed that it performs better than other methods, especially when we only have a small amount of data. |
Keywords
» Artificial intelligence » Instruction tuning » Language model
