Timezone: »
Beyond Scale: the Diversity Coefficient as a Data Quality Metric Demonstrates LLMs are Pre-trained on Formally Diverse Data
Alycia Lee · Brando Miranda · Sanmi Koyejo
Event URL: https://openreview.net/forum?id=oCYjN48axE »
Current trends to pre-train capable Large Language Models (LLMs) mostly focus on scaling of model and dataset size.However, the $\textit{quality}$ of pre-training data is an important factor for training powerful LLMs, yet it is a nebulous concept that has not been fully characterized.Therefore, we use the recently proposed Task2Vec diversity coefficient to understand formal aspects of data quality that go beyond scale alone.Specifically, we measure the diversity coefficient of publicly available pre-training datasets to demonstrate that their formal diversity is high when compared to theoretical lower and upper bounds.In addition, to build confidence in the diversity coefficient, we conduct interpretability experiments and find that the coefficient aligns with intuitive properties of diversity,e.g., it increases as the number of latent concepts increases. We conclude the diversity coefficient is reliable and conjecture it can be used to build useful diverse datasets for LLMs.
Current trends to pre-train capable Large Language Models (LLMs) mostly focus on scaling of model and dataset size.However, the $\textit{quality}$ of pre-training data is an important factor for training powerful LLMs, yet it is a nebulous concept that has not been fully characterized.Therefore, we use the recently proposed Task2Vec diversity coefficient to understand formal aspects of data quality that go beyond scale alone.Specifically, we measure the diversity coefficient of publicly available pre-training datasets to demonstrate that their formal diversity is high when compared to theoretical lower and upper bounds.In addition, to build confidence in the diversity coefficient, we conduct interpretability experiments and find that the coefficient aligns with intuitive properties of diversity,e.g., it increases as the number of latent concepts increases. We conclude the diversity coefficient is reliable and conjecture it can be used to build useful diverse datasets for LLMs.
Author Information
Alycia Lee (Stanford University)
Brando Miranda (Stanford University)
Sanmi Koyejo (Stanford University)
More from the Same Authors
-
2023 : Layer-Wise Feedback Alignment is Conserved in Deep Neural Networks »
Zach Robertson · Sanmi Koyejo -
2023 : FACADE: A Framework for Adversarial Circuit Anomaly Detection and Evaluation »
Dhruv Pai · Andres Carranza · Rylan Schaeffer · Arnuv Tandon · Sanmi Koyejo -
2023 : Beyond Scale: the Diversity Coefficient as a Data Quality Metric Demonstrates LLMs are Pre-trained on Formally Diverse Data »
Alycia Lee · Brando Miranda · Brando Miranda · Sanmi Koyejo -
2023 : Leveraging Side Information for Communication-Efficient Federated Learning »
Berivan Isik · Francesco Pase · Deniz Gunduz · Sanmi Koyejo · Tsachy Weissman · Michele Zorzi -
2023 : GPT-Zip: Deep Compression of Finetuned Large Language Models »
Berivan Isik · Hermann Kumbong · Wanyi Ning · Xiaozhe Yao · Sanmi Koyejo · Ce Zhang -
2023 : Are Emergent Abilities of Large Language Models a Mirage? »
Rylan Schaeffer · Brando Miranda · Sanmi Koyejo -
2023 : Thomas: Learning to Explore Human Preference via Probabilistic Reward Model »
Sang Truong · Duc Nguyen · Tho Quan · Sanmi Koyejo -
2023 : On learning domain general predictors »
Sanmi Koyejo -
2023 : Deceptive Alignment Monitoring »
Andres Carranza · Dhruv Pai · Rylan Schaeffer · Arnuv Tandon · Sanmi Koyejo -
2023 : Vignettes on Pairwise-Feedback Mechanisms for Learning with Uncertain Preferences »
Sanmi Koyejo -
2023 Poster: Pairwise Ranking Losses of Click-Through Rates Prediction for Welfare Maximization in Ad Auctions »
Boxiang Lyu · Zhe Feng · Zach Robertson · Sanmi Koyejo