Predicting Data Efficiency for LLM Fine-tuning
Published:Dec 31, 2025 17:37
•1 min read
•ArXiv
Analysis
This paper addresses the practical problem of determining how much data is needed to fine-tune large language models (LLMs) effectively. It's important because fine-tuning is often necessary to achieve good performance on specific tasks, but the amount of data required (data efficiency) varies greatly. The paper proposes a method to predict data efficiency without the costly process of incremental annotation and retraining, potentially saving significant resources.
Key Takeaways
- •Addresses the problem of unknown data efficiency in LLM fine-tuning.
- •Proposes a method to predict data efficiency using gradient cosine similarity.
- •Aims to reduce the need for costly incremental annotation and retraining.
- •Achieves 8.6% error in data efficiency prediction on a diverse set of tasks.
Reference
“The paper proposes using the gradient cosine similarity of low-confidence examples to predict data efficiency based on a small number of labeled samples.”