Text Quality-Based Pruning for Efficient Training of Language Models
Vasu Sharma, Karthik Padthe, Newsha Ardalani, Kushal Tirumala, Russell Howes, Hu Xu, Po-Yao Huang, Shang-Wen Li, Armen Aghajanyan, Gargi Ghosh, Luke Zettlemoyer
TL;DR
The paper tackles the inefficiency of training language models on large, noisy corpora by introducing a model-agnostic text quality score derived from 14 heuristic filters and perplexity-based weighting. It computes line-level scores with $score_{line}=\frac{\sum_{i=1}^{F} w_i I_i(line)}{\sum_{i=1}^{F} w_i}$ and aggregates to $score_{doc}=\frac{\sum_{line} tc_{line} score_{line}}{\sum_{line} tc_{line}}$, where $w_i=\max(0, (PPL_{all}-PPL_i)/PPL_{all})$, then prunes data by keeping high-quality samples in top percentiles. Empirical results on OpenWebText and Wikipedia show that pruning around 20-40% of data can yield up to 0.9% absolute accuracy gains on 14 tasks while training faster by about 21-42%, across several architectures (GPT-2, GPT-Neo, Pythia, OPT-125M). The framework demonstrates a practical, scalable approach to dataset curation in LM training and outlines limitations for larger models and datasets, proposing directions for future work.
Abstract
In recent times training Language Models (LMs) have relied on computationally heavy training over massive datasets which makes this training process extremely laborious. In this paper we propose a novel method for numerically evaluating text quality in large unlabelled NLP datasets in a model agnostic manner to assign the text instances a "quality score". By proposing the text quality metric, the paper establishes a framework to identify and eliminate low-quality text instances, leading to improved training efficiency for LM models. Experimental results over multiple models and datasets demonstrate the efficacy of this approach, showcasing substantial gains in training effectiveness and highlighting the potential for resource-efficient LM training. For example, we observe an absolute accuracy improvement of 0.9% averaged over 14 downstream evaluation tasks for multiple LM models while using 40% lesser data and training 42% faster when training on the OpenWebText dataset and 0.8% average absolute accuracy improvement while using 20% lesser data and training 21% faster on the Wikipedia dataset.