Rethinking data selection: coverage matters more than difficulty in LLM fine-tuning

Problem

To cut fine-tuning costs, you need to select and use only part of the training data. Earlier work built various difficulty scores on the assumption that 'harder examples are more useful.' But these scores were validated mainly on small classification models. It was unclear whether they also work for LLM fine-tuning, where the output space is the entire vocabulary.

Method

The authors adapted five difficulty scores for LLM fine-tuning: Perplexity, Confidence, Variability, VoG (Variance of Gradients) and PVI. They fine-tuned Llama 3 8B and OLMo 2 7B on Social IQa, CommonsenseQA and CosmosQA using 1–75% of the training data and compared the results with random selection. Coverage was measured as follows. The last-layer Hidden Representations of the model before fine-tuning are clustered with k-means. Then the Jensen-Shannon Divergence (JSD) between the cluster distributions of the full dataset and the selected subset is computed. Finally, the authors implemented a simple clustering-based selection method that picks data representative of the input distribution.

Results

• Across all three datasets and both models, random selection matched or often beat the five difficulty-based methods. Among the difficulty methods, VoG had the highest average accuracy.
• Using only 50% of the data gave nearly the same performance as training on 100%. The authors interpreted this as evidence that the datasets contain a lot of redundancy.
• Across all datasets, models and selection ratios, random selection had the lowest average JSD (the highest coverage). The Spearman correlation between average JSD and accuracy was −0.71 to −0.90, a strong negative correlation.
• The authors report that clustering-based selection matched or exceeded random selection on single-domain tasks. Performance increased monotonically as the selection ratio grew, which also avoided the instability of random selection.
• On the Tülu 3 SFT data, which mixes many domains, the authors report that clustering-based selection outperformed all compared methods, especially at small data ratios.

Limitations

The main experiments are limited to two 7–8B models and three multiple-choice QA datasets. Evaluation uses first-token exact match. In single-domain settings, the proposed method is roughly on par with random selection and only occasionally ahead. The coverage measure is sensitive to the choice of k and embeddings, so results were averaged over several k values and seeds. The provided text cuts off partway through the description of the clustering method, so detailed figures for the proposed method are summarized only from the abstract and introduction.

Why it matters

When reducing fine-tuning data, random selection or selection that evenly covers the distribution may be a safer default than elaborate difficulty scores. In data that mixes several domains in particular, random selection alone can easily lose balance. In such cases, selection based on embedding clusters is a practical alternative.

Key terms

Data Selection

A technique that trains on only a selected part of the training data to cut costs while keeping performance.

Difficulty Score

A score that rates how hard each example is, using signals such as the model's loss, confidence or changes in gradients. Harder examples are usually selected first.

Coverage

How evenly the selected data represents the full input distribution. This paper measured it as the difference between distributions of embedding clusters.

Jensen-Shannon Divergence(JSD)

A symmetric measure of how different two probability distributions are. A lower value means the selected data is closer to the full distribution.

Original

• arXiv abstract
• PDF
• HTML full text
• Hugging Face

Authors: Lalchand Pandia, Kanishka Misra, Allyson Ettinger · Posted 2026-11-05 · License arXiv 비독점 배포 라이선스 · Based on the full text

This review is an AI-written summary of the paper. Please treat the original paper as the reference.