The Unreasonable Effectiveness of Easy Training Data for Hard Tasks

2024/01/12 by Peter Hase, Hase, Peter, Mohit Bansal +5 · 10 citations
Computer Science · #Artificial Intelligence (cs.AI) #Artificial intelligence #Classifier (UML) #Code (set theory) #Computation and Language (cs.CL) #Computer science #Context (archaeology) #Data science #Database #FOS: Computer and information sciences #Generalization #Language model #Machine Learning (cs.LG) #Machine learning #Natural Language Processing Techniques #Natural language processing #Oracle #Programming language #Scalability #Set (abstract data type) #Simple (philosophy) #Test data #Text Readability and Simplification #Topic Modeling #Training set

paper · pdf · doi:10.48550/arxiv.2401.06751

published in arXiv (Cornell University) (Cornell University)

openalex publication_date 2024/01/12 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28

Abstract

How can we train models to perform well on hard test data when hard training data is by definition difficult to label correctly? This question has been termed the scalable oversight problem and has drawn increasing attention as language models have continually improved. In this paper, we present the surprising conclusion that current pretrained language models often generalize relatively well from easy to hard data, even performing as well as oracle models finetuned on hard data. We demonstrate this kind of easy-to-hard generalization using simple finetuning methods like in-context learning, linear classifier heads, and QLoRA for seven different measures of datapoint hardness, including six empirically diverse human hardness measures (like grade level) and one model-based measure (loss-based). Furthermore, we show that even if one cares most about model performance on hard data, it can be better to collect easy data rather than hard data for finetuning, since hard data is generally noisier and costlier to collect. Our experiments use open models up to 70b in size and four publicly available question-answering datasets with questions ranging in difficulty from 3rd grade science questions to college level STEM questions and general-knowledge trivia. We conclude that easy-to-hard generalization in LMs is surprisingly strong for the tasks studied. Our code is available at: https://github.com/allenai/easy-to-hard-generalization

Cited by

• Pre-Generating Multi-Difficulty PDE Data for Few-Shot Neural PDE Solvers
• Revisiting Generalization Across Difficulty Levels: It's Not So Easy
• ACE and Diverse Generalization via Selective Disagreement
• FLAMES: Improving LLM Math Reasoning via a Fine-Grained Analysis of the Data Synthesis Pipeline
• ReSURE: Regularizing Supervision Unreliability for Multi-turn Dialogue Fine-tuning
• The Delta Learning Hypothesis: Preference Tuning on Weak Data can Yield Strong Gains
• Do Language Models Mirror Human Confidence? Exploring Psychological Insights to Address Overconfidence in LLMs
• Learning Composable Chains-of-Thought
• EnsemW2S: Enhancing Weak-to-Strong Generalization with Large Language Model Ensembles
• Alice: Proactive Learning with Teacher's Demonstrations for Weak-to-Strong Generalization

Related

• Revisiting Generalization Across Difficulty Levels: It's Not So Easy
• Can Large Language Models Always Solve Easy Problems if They Can Solve Harder Ones?
• Zero-to-Strong Generalization: Eliciting Strong Capabilities of Large Language Models Iteratively without Gold Labels
• LLMs Encode How Difficult Problems Are
• Quantifying Generalization Complexity for Large Language Models