plawanrath/gemma-2-9b-it-wanda-s50-pia
The plawanrath/gemma-2-9b-it-wanda-s50-pia model is a 9 billion parameter instruction-tuned Gemma-2 variant, developed by Plawan Kumar Rath, that has undergone 50% target Wanda pruning. 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 documented bias amplification and lack of inference or storage benefits from unstructured pruning.
Loading preview...
Overview
This model, gemma-2-9b-it-wanda-s50-pia, is a 9 billion parameter instruction-tuned Gemma-2 variant created by Plawan Kumar Rath. It is a research artifact specifically designed to study the impact of weight pruning on fairness in large language models. The model was pruned using the wanda method with a target sparsity of 50%, achieving an actual sparsity of 33.33% by zeroing approximately 2.77 billion parameters in the linear layers of transformer blocks.
Key Findings & Caveats
- Bias Amplification: The primary finding from the associated paper, "Weight Pruning Amplifies Bias: A Multi-Method Study of Compressed LLMs for Edge AI" (IEEE AIIoT 2026), is that Wanda pruning at this sparsity level induces measurable bias amplification on the BBQ benchmark. This is referred to as the "Smart Pruning Paradox," where perplexity-based evaluations may not reveal significant bias increases.
- No Performance Benefits: Despite pruning, this unstructured pruning method provides no storage savings (on-disk size is identical to the dense base model) and no inference latency savings (inference time is identical to the dense baseline on common hardware like Apple Silicon).
- Research Use Only: Due to the documented bias amplification and lack of practical deployment benefits, this model is explicitly not for production use in any user-facing or decision-making system.
Good for
- Academic Research: Studying the effects of weight pruning on model fairness and bias amplification.
- Understanding Pruning Limitations: Investigating why unstructured pruning may not yield expected benefits in storage or latency on current hardware.
- Reproducing Research: Replicating the findings presented in the companion paper by Plawan Kumar Rath and Rahul Maliakkal.