OptiPFair: A Structured Pruning and Bias Visualization Tool for Large Language Models

OptiPFair is a Python tool library designed for structured pruning and fairness analysis of large language models (LLMs). It supports activation value analysis, bias detection and mitigation, helping developers compress models while identifying and reducing potential biases.

Basic project info:

• Author/Maintainer: Pere Martra (peremartra)
• Source: GitHub (repo link: https://github.com/peremartra/optipfair; project homepage: https://peremartra.github.io/optipfair/)
• License: Apache License 2.0
• Creation time: April 11, 2025; Last update: May 30, 2026

With the continuous growth of LLM scale, reducing computational costs while maintaining performance has become a key challenge. Structured pruning (removing entire neurons or attention heads) is suitable for hardware acceleration but may introduce or amplify model biases, affecting fairness across different groups. OptiPFair was created to address this dual challenge: optimizing model efficiency and ensuring fairness.

Structured Pruning

OptiPFair implements structured pruning methods for LLMs, including:

• Attention head pruning: Remove low-impact heads to reduce computational complexity.
• Neuron pruning: Remove redundant neurons based on activation analysis.
• Hierarchical pruning: Flexible granularity from global to local. Notably, pruning decisions consider both performance and fairness constraints.

Bias Detection & Visualization

• Activation analysis: Monitor internal activation patterns to identify neurons with systemic bias against specific groups.
• Bias heatmaps: Visualize bias distribution across different inputs.
• Fairness metrics: Calculate indicators like demographic parity and equal opportunity.

Technical Implementation

• Based on Python, deeply integrated with Hugging Face Transformers.
• Uses PyTorch's automatic differentiation to preserve gradient flow during pruning.
• Modular design: Pruning, bias detection, and visualization components can be used independently or combined.
• Compatible with mainstream Transformer architectures (GPT, BERT, T5, etc.)

1. Edge Device Deployment: Compress model size by 30%-60% for resource-constrained environments (mobile/embedded systems) while monitoring and mitigating pruning-induced biases.
2. Fairness Audit: Help research teams and enterprises audit existing models for biases in sensitive attributes (gender, race, age) to support model improvement.
3. Model Optimization Iteration: Evaluate the impact of different architectures/training strategies on both efficiency and fairness during model iteration.

• Community recognition: 40 GitHub stars and 9 forks.
• Tags: Activation analysis, bias detection, fairness, LLM, model pruning (aligned with AI ethics and efficiency research hotspots).
• Documentation: Detailed docs (https://peremartra.github.io/optipfair/) including API references, tutorials, and example code, lowering the entry barrier.

OptiPFair represents an important direction for LLM optimization tools: balancing efficiency and fairness instead of focusing solely on performance. This dual-objective approach is crucial for responsible AI development.

For developers and researchers aiming to deploy efficient and fair LLMs, OptiPFair is a valuable tool. As model scales grow and AI ethics regulations improve, such tools will play an increasingly important role in the model development lifecycle.