Quantifying the Importance of Data Alignment in Downstream Model Performance

TL;DR

This paper demonstrates that data alignment between training and evaluation datasets significantly impacts large language model performance, often more than dataset size, especially in specialized tasks like Autoformalization.

Contribution

It introduces a quantitative measure of data alignment and shows its strong correlation with downstream performance, challenging the focus on data quantity in LLM training.

Findings

Higher data alignment correlates with lower model loss.

Data alignment impacts performance more than dataset size.

Alignment measurement guides better training strategies.

Abstract

Contrary to the conventional emphasis on dataset size, we explore the role of data alignment -- an often overlooked aspect of data quality -- in training capable Large Language Models (LLMs). To do so, we use the Task2Vec-based alignment coefficient, a quantitative measure of the similarity between two datasets, to quantify the impact of alignment between training data and evaluation data on downstream performance. In particular, we conduct controlled \textit{interventional} experiments for two settings: 1. the impact of increased alignment coefficients between various pre-training (pt) against evaluation datasets, and 2. the impact of increased alignment coefficients between domain specific fine-tuning (ft) against domain specific evaluation. The domain specific task we explore is Autoformalization -- the machine translation task between natural language and code for formal verification. In both settings, we find a strong, predictable negative correlation between the alignment coefficient of a model's training and evaluation data and the model's loss/perplexity on the respective downstream task. These findings suggest a re-evaluation of LLM training approaches, demonstrating the relevance of data alignment compared to data quantity, especially in specialized downstream tasks such as Autoformalization.