Summary of Automatic Pruning Of Fine-tuning Datasets For Transformer-based Language Models, by Mohammadreza Tayaranian et al.
Automatic Pruning of Fine-tuning Datasets for Transformer-based Language Models
by Mohammadreza Tayaranian, Seyyed Hasan Mozafari, Brett H. Meyer, James J. Clark, Warren J. Gross
First submitted to arxiv on: 11 Jul 2024
Categories
- Main: Computation and Language (cs.CL)
- Secondary: 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 The proposed automatic dataset pruning method for fine-tuning tasks leverages transformer-based language models to achieve state-of-the-art performance on natural language understanding tasks. By pre-training on general corpus and then fine-tuning, these models excel on downstream tasks. The study investigates the effect of pruning the training set on the evaluation set’s performance. Instead of relying on user feedback, this method automatically extracts subsets for each pair of model and fine-tuning task based on the model’s success rate in correctly classifying data points. This approach provides multiple subsets to navigate the trade-off between subset size and evaluation accuracy. The winning ticket subset is significantly smaller than the original training set (averaging 3x reduction) and, on average, fine-tuning on these subsets boosts evaluation performance by 0.1%. Experiments on 5 downstream tasks and 2 language models validate this method’s effectiveness. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Transformer-based language models have improved natural language understanding tasks. This study explores pruning the training set for fine-tuning tasks to achieve better results. The new method automatically selects subsets based on how well the model does on each data point. This approach helps balance the trade-off between fewer training examples and better performance. The best subset, which is 3 times smaller than usual, makes models perform 0.1% better. |
Keywords
* Artificial intelligence * Fine tuning * Language understanding * Pruning * Transformer
