🤖 AI Summary
Current supervised fine-tuning (SFT) lacks efficient and interpretable data selection methods; blindly scaling data volume often fails to improve performance while incurring high preprocessing and validation costs. This paper proposes KAMIR, a data selection framework that evaluates sample familiarity based on internal model representations. Its core innovation lies in quantifying knowledge familiarity via layer-wise hidden-state similarity to the final output state—eliminating reliance on prompt engineering. A lightweight classifier identifies low-familiarity samples, enabling task-agnostic, few-shot-friendly training data filtering. Evaluated across diverse tasks—including machine reading comprehension and summarization—KAMIR achieves significant generalization gains using only a small subset of low-familiarity samples for fine-tuning. Empirical results validate the efficacy of the “less-but-better” data curation paradigm, demonstrating that targeted selection of semantically challenging instances yields superior performance over brute-force data expansion.
📝 Abstract
Recent advances in large language models (LLMs) have been driven by pretraining, supervised fine tuning (SFT), and alignment tuning. Among these, SFT plays a crucial role in transforming a model 's general knowledge into structured responses tailored to specific tasks. However, there is no clearly established methodology for effective training data selection. Simply increasing the volume of data does not guarantee performance improvements, while preprocessing, sampling, and validation require substantial time and cost.
To address this issue, a variety of data selection methods have been proposed. Among them, knowledge based selection approaches identify suitable training data by analyzing the model 's responses. Nevertheless, these methods typically rely on prompt engineering, making them sensitive to variations and incurring additional costs for prompt design.
In this study, we propose Knowledge Analysis via Model Internal Representations (KAMIR), a novel approach that overcomes these limitations by analyzing data based on the model 's internal representations. KAMIR computes similarities between the hidden states of each layer (block) and the final hidden states for a given input to assess the data. Unlike prior methods that were largely limited to multiple choice tasks, KAMIR can be applied to a wide range of tasks such as machine reading comprehension and summarization. Moreover, it selects data useful for training based on the model 's familiarity with the input, even with a small dataset and a simple classifier architecture. Experiments across diverse task datasets demonstrate that training with less familiar data leads to better generalization performance.