plawanrath/gemma-2-9b-it-magnitude-s70-pia
The plawanrath/gemma-2-9b-it-magnitude-s70-pia is a 9 billion parameter instruction-tuned language model, derived from Google's Gemma-2-9B-IT, that has undergone magnitude-based unstructured pruning to achieve 70% sparsity. This model serves as a research artifact to study fairness degradation under weight pruning, specifically demonstrating bias amplification on the BBQ benchmark. It is not intended for production use due to its research-focused nature and identified bias issues, and notably offers no storage or latency savings over its dense baseline.
Loading preview...
Overview of plawanrath/gemma-2-9b-it-magnitude-s70-pia
This model is a 9 billion parameter instruction-tuned variant of Google's Gemma-2-9B-IT, specifically created as a research artifact to investigate the impact of weight pruning on fairness in large language models. It has been subjected to magnitude-based unstructured pruning, achieving a 70% sparsity level (70.13% actual sparsity) by zeroing weights with the smallest absolute values in the transformer block's linear layers.
Key Research Findings & Characteristics
- Bias Amplification: The primary finding from the companion paper "Weight Pruning Amplifies Bias: A Multi-Method Study of Compressed LLMs for Edge AI" (IEEE AIIoT 2026) is that this level of pruning induces measurable bias amplification on the BBQ benchmark. This highlights a critical issue where compression methods can exacerbate existing biases.
- No Performance Gains: Despite significant pruning, the model exhibits no storage savings (on-disk size is identical to the dense baseline as unstructured sparsity is not exploited by formats like SafeTensors/GGUF) and no inference latency savings (identical to the dense baseline on Apple Silicon/MLX, as dense GEMM kernels do not skip zero entries).
- Research-Only: This model is explicitly designated for research purposes only and is not suitable for production deployment in user-facing or decision-making systems due to its demonstrated bias amplification.
- Pruning Scope: Pruning was applied to linear layers within transformer blocks (attention projections + MLP), while embeddings, LM head, and layer norms remained untouched.
Good for
- Academic Research: Studying the effects of weight pruning on model fairness and bias amplification.
- Understanding Pruning Limitations: Investigating scenarios where unstructured pruning does not yield expected storage or latency benefits.
- Reproducibility: Replicating the findings presented in the associated research paper on pruning impact analysis.