Improving Language Models through Interventions

Introduction

• Language Models (LMs):

  • Used in various fields such as medicine, finance, science, and entertainment.
  • Can behave unpredictably, generate incorrect or harmful content.
  • User needs change due to new regulations, limited resources, outdated knowledge, or copyright issues.
  • Need for quick solutions to prevent LMs from becoming inaccurate, outdated, or biased.

• Interventions:

  • Efficient updates for LMs after initial training.
  • Aim to address issues such as outdated information and changing user requirements.
  • Examples: model compression, knowledge editing, and unlearning.
  • Typically developed independently, posing challenges for effective combination.

• Proposed Solution:

  • Interventions should be composable, meaning they can work together harmoniously.
  • Metrics introduced:
    • Order-free error: Assesses if one intervention impacts the success of others.
    • Order sensitivity: Checks if the application order of interventions affects the outcome.

• Experiments:

  • Conducted using the LLaMA 3-8B model.
  • Results showed that model compression can hinder other interventions.
  • Emphasis on developing interventions with composability in mind.

Composable Interventions for Language Models

• Intervention Composition:

  • Involves applying multiple interventions in a specific order to a model.
  • Assume each intervention is based on a single criterion for simplicity.

• Measuring Effectiveness:

  • Consider how one intervention affects the application of subsequent interventions.
  • Metrics introduced:
    • Order-free error: Performance regardless of order.
    • Order sensitivity: Performance invariance to order.

• Evaluation Method:

  • Apply an intervention, measure its impact, add another, and measure the new impact.
  • Reverse the order of interventions and compare results.
  • Helps identify direct interactions between interventions.

• Experimental Setup:

  • Study of the impact of multiple sequential interventions on model performance.
  • Importance of order in applying interventions for effective results.
  • Detailed evaluation crucial for developing practical and composable interventions.

Implementation Details

• Model & Metrics:

  • Experiments conducted with the LLaMA 3-8B model.
  • Focus on how interventions work individually and combined.

• Findings:

  • Model compression impacts the success of knowledge editing and overall performance.
  • Order of applying interventions affects the outcome.
  • Emphasis on creating tailored methods for successful composition.

• Utility Evaluations:

  • Overall utility evaluations may not accurately capture composability.
  • Need for detailed evaluations focusing on composability as a design factor.

Composing Model Compression with Machine Unlearning

• Interventions:

  • Composed three model compression and three machine unlearning interventions.
  • Evaluated success of unlearning using wdp (weighted data points).

• Results:

  • Compressing models can make unlearning more difficult.
  • Pruning before unlearning reduces performance, especially at higher sparsity.
  • Optimal order varies depending on the method and intervention.
  • Compression tends to hinder other interventions and make targeted updates harder.

• Metrics & Techniques:

  • GD and RMU outperform Gau in terms of order-free error on wdp.
  • RMU shows lower overall order sensitivity, while GD has higher order sensitivity for pruning compared to quantization.

• Knowledge Editing and Unlearning:

  • High composability between some unlearning methods and knowledge editing.
  • RMU stands out due to low order-free error and order sensitivity.
  • Editing precise modifications do not disrupt unlearning targets.

• Summary:

  • Compression alters model knowledge storage, making updates difficult.
  • RMU emerges as the most composable unlearning technique.
  • Need thorough evaluations using multiple metrics and datasets to assess intervention composability effectively.