Does your data spark joy? Performance gains from domain upsampling at the end of training

• URL: http://arxiv.org/abs/2406.03476v1
• Date: Wed, 5 Jun 2024 17:29:15 GMT
• Title: Does your data spark joy? Performance gains from domain upsampling at the end of training
• Authors: Cody Blakeney, Mansheej Paul, Brett W. Larsen, Sean Owen, Jonathan Frankle,
• Abstract summary: It is expensive to understand the impact of domain-specific datasets on training at large FL model scales. We use domain upsampling to characterize at scale the utility of individual datasets for improving various benchmarks. This tool opens up the ability to experiment with the impact of different pretraining datasets at scale, but at an order of lower cost compared to full pretraining runs.
• Score: 16.572129046599937
• License: http://creativecommons.org/licenses/by-nc-sa/4.0/
• Abstract: Pretraining datasets for large language models (LLMs) have grown to trillions of tokens composed of large amounts of CommonCrawl (CC) web scrape along with smaller, domain-specific datasets. It is expensive to understand the impact of these domain-specific datasets on model capabilities as training at large FLOP scales is required to reveal significant changes to difficult and emergent benchmarks. Given the increasing cost of experimenting with pretraining data, how does one determine the optimal balance between the diversity in general web scrapes and the information density of domain specific data? In this work, we show how to leverage the smaller domain specific datasets by upsampling them relative to CC at the end of training to drive performance improvements on difficult benchmarks. This simple technique allows us to improve up to 6.90 pp on MMLU, 8.26 pp on GSM8K, and 6.17 pp on HumanEval relative to the base data mix for a 7B model trained for 1 trillion (T) tokens, thus rivaling Llama-2 (7B)$\unicode{x2014}$a model trained for twice as long. We experiment with ablating the duration of domain upsampling from 5% to 30% of training and find that 10% to 20% percent is optimal for navigating the tradeoff between general language modeling capabilities and targeted benchmarks. We also use domain upsampling to characterize at scale the utility of individual datasets for improving various benchmarks by removing them during this final phase of training. This tool opens up the ability to experiment with the impact of different pretraining datasets at scale, but at an order of magnitude lower cost compared to full pretraining runs.

Related papers

• AutoScale: Automatic Prediction of Compute-optimal Data Composition for Training LLMs [61.13296177652599]
  This paper demonstrates that the optimal composition of training data from different domains is scale-dependent. We introduce *AutoScale*, a novel, practical approach for optimizing data compositions at potentially large training data scales. Our evaluation on GPT-2 Large and BERT pre-training demonstrates *AutoScale*'s effectiveness in improving training convergence and downstream performance.
  arXiv  Detail & Related papers  (2024-07-29T17:06:30Z)
• Effective pruning of web-scale datasets based on complexity of concept clusters [48.125618324485195]
  We present a method for pruning large-scale multimodal datasets for training CLIP-style models on ImageNet. We find that training on a smaller set of high-quality data can lead to higher performance with significantly lower training costs. We achieve a new state-of-the-art Imagehttps://info.arxiv.org/help/prep#commentsNet zero-shot accuracy and a competitive average zero-shot accuracy on 38 evaluation tasks.
  arXiv  Detail & Related papers  (2024-01-09T14:32:24Z)
• Efficient Grammatical Error Correction Via Multi-Task Training and Optimized Training Schedule [55.08778142798106]
  We propose auxiliary tasks that exploit the alignment between the original and corrected sentences. We formulate each task as a sequence-to-sequence problem and perform multi-task training. We find that the order of datasets used for training and even individual instances within a dataset may have important effects on the final performance.
  arXiv  Detail & Related papers  (2023-11-20T14:50:12Z)
• When Less is More: Investigating Data Pruning for Pretraining LLMs at Scale [12.94829977468838]
  Large volumes of text data have contributed significantly to the development of large language models. To date, efforts to prune datasets down to a higher quality subset have relied on hand-crafteds encoded as rule-based filters. We take a wider view and explore scalable estimates of data quality that can be used to measure the quality of pretraining data.
  arXiv  Detail & Related papers  (2023-09-08T19:34:05Z)
• D4: Improving LLM Pretraining via Document De-Duplication and Diversification [38.84592304799403]
  We show that careful data selection via pre-trained model embeddings can speed up training. We also show that repeating data intelligently consistently outperforms baseline training.
  arXiv  Detail & Related papers  (2023-08-23T17:58:14Z)
• Delving Deeper into Data Scaling in Masked Image Modeling [145.36501330782357]
  We conduct an empirical study on the scaling capability of masked image modeling (MIM) methods for visual recognition. Specifically, we utilize the web-collected Coyo-700M dataset. Our goal is to investigate how the performance changes on downstream tasks when scaling with different sizes of data and models.
  arXiv  Detail & Related papers  (2023-05-24T15:33:46Z)
• DoReMi: Optimizing Data Mixtures Speeds Up Language Model Pretraining [148.90031913522648]
  We propose Domain Reweighting with Minimax Optimization (DoReMi) DoReMi first trains a small proxy model using group distributionally robust optimization (Group DRO) over domains to produce domain weights. We then resample a dataset with these domain weights and train a larger, full-sized model.
  arXiv  Detail & Related papers  (2023-05-17T17:58:13Z)
• Dataset Pruning: Reducing Training Data by Examining Generalization Influence [30.30255670341501]
  Do all training data contribute to model's performance? How to construct a smallest subset from the entire training data as a proxy training set without significantly sacrificing the model's performance?
  arXiv  Detail & Related papers  (2022-05-19T05:36:35Z)
• Self-Supervised Pre-Training for Transformer-Based Person Re-Identification [54.55281692768765]
  Transformer-based supervised pre-training achieves great performance in person re-identification (ReID) Due to the domain gap between ImageNet and ReID datasets, it usually needs a larger pre-training dataset to boost the performance. This work aims to mitigate the gap between the pre-training and ReID datasets from the perspective of data and model structure.
  arXiv  Detail & Related papers  (2021-11-23T18:59:08Z)
• Omni-supervised Facial Expression Recognition via Distilled Data [120.11782405714234]
  We propose omni-supervised learning to exploit reliable samples in a large amount of unlabeled data for network training. We experimentally verify that the new dataset can significantly improve the ability of the learned FER model. To tackle this, we propose to apply a dataset distillation strategy to compress the created dataset into several informative class-wise images.
  arXiv  Detail & Related papers  (2020-05-18T09:36:51Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.