Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Beyond Perplexity: A Lightweight Benchmark for Knowledge Retention in Supervised Fine-Tuning

Soheil Zibakhsh Shabgahi, Pedram Aghazadeh, Farinaz Koushanfar

TL;DR

The paper addresses the inadequacy of validation perplexity for tracking factual knowledge during supervised fine-tuning of LLMs. It introduces the Knowledge Retention Test (KR-Test), a lightweight, corpus-grounded, contrastive evaluation that uses semantic segmentation, contrastive generation, and likelihood assessment to distinguish factual learning from linguistic mimicry, without instruction tuning or decoding. An oracle-based validation confirms the benchmark's discriminative power, and KR-Test is used to dissect PEFT dynamics, notably LoRA placement (FFN vs Attention) and model capacity effects, revealing a dissociation between linguistic convergence and knowledge retention. The framework enables a practical, compute-efficient signal for early stopping, hyperparameter tuning, and informed PEFT design, with broader implications for knowledge scaling laws in language models. KR-Test benefits from open-source curation and provides a principled path to interpret fine-tuning dynamics beyond perplexity.

Abstract

Supervised Fine-Tuning (SFT) is a standard approach for injecting domain knowledge into Large Language Models (LLMs). However, relying on validation perplexity to monitor training is often insufficient, as it confounds stylistic mimicry with genuine factual internalization. To address this, we introduce the Knowledge Retention (KR) Test , a lightweight, corpus-grounded evaluation framework designed to distinguish factual learning from linguistics. KR-Test utilizes automatically generated contrastive examples to measure likelihood preferences for correct versus incorrect continuations, requiring no instruction tuning or generative decoding. We validate the framework's integrity through a "blind vs. oracle" baseline analysis. Furthermore, we demonstrate the diagnostic capabilities of KR-Test by analyzing the training dynamics of Low-Rank Adaptation (LoRA). By exposing the fine-grained dissociation between linguistic convergence and knowledge retention, KR-Test enhances the interpretability of fine-tuning dynamics.

Beyond Perplexity: A Lightweight Benchmark for Knowledge Retention in Supervised Fine-Tuning

TL;DR

The paper addresses the inadequacy of validation perplexity for tracking factual knowledge during supervised fine-tuning of LLMs. It introduces the Knowledge Retention Test (KR-Test), a lightweight, corpus-grounded, contrastive evaluation that uses semantic segmentation, contrastive generation, and likelihood assessment to distinguish factual learning from linguistic mimicry, without instruction tuning or decoding. An oracle-based validation confirms the benchmark's discriminative power, and KR-Test is used to dissect PEFT dynamics, notably LoRA placement (FFN vs Attention) and model capacity effects, revealing a dissociation between linguistic convergence and knowledge retention. The framework enables a practical, compute-efficient signal for early stopping, hyperparameter tuning, and informed PEFT design, with broader implications for knowledge scaling laws in language models. KR-Test benefits from open-source curation and provides a principled path to interpret fine-tuning dynamics beyond perplexity.

Abstract

Supervised Fine-Tuning (SFT) is a standard approach for injecting domain knowledge into Large Language Models (LLMs). However, relying on validation perplexity to monitor training is often insufficient, as it confounds stylistic mimicry with genuine factual internalization. To address this, we introduce the Knowledge Retention (KR) Test , a lightweight, corpus-grounded evaluation framework designed to distinguish factual learning from linguistics. KR-Test utilizes automatically generated contrastive examples to measure likelihood preferences for correct versus incorrect continuations, requiring no instruction tuning or generative decoding. We validate the framework's integrity through a "blind vs. oracle" baseline analysis. Furthermore, we demonstrate the diagnostic capabilities of KR-Test by analyzing the training dynamics of Low-Rank Adaptation (LoRA). By exposing the fine-grained dissociation between linguistic convergence and knowledge retention, KR-Test enhances the interpretability of fine-tuning dynamics.
Paper Structure (26 sections, 1 equation, 4 figures)

This paper contains 26 sections, 1 equation, 4 figures.

Table of Contents

  1. Introduction
  2. Contributions.
  3. Limitations of Perplexity for Factual Evaluation
  4. Knowledge Retention Test
  5. Semantic Segmentation.
  6. Contrastive Generation.
  7. Likelihood Evaluation.
  8. Test Curation
  9. Experiments
  10. Oracle-Based Validation
  11. Demystifying PEFT Dynamics with KR-Test
  12. Setup
  13. Results and Analysis
  14. Implications for Model Selection
  15. Capacity and Knowledge Scaling
  16. ...and 11 more sections

Figures (4)

  • Figure 1: KR-Test generation and validation pipeline. Questions are generated using segmented data and are used in validation to track the model's learning progress.
  • Figure 2: Parameter efficiency of LoRA module placement according to KR-Test.
  • Figure 3: Effect of parameter count on model's initial and final knowledge retention
  • Figure A1: Example KR-Test instance derived from WikiText illustrating minimal factual perturbation.