plawanrath/mistral-7b-instruct-v0.3-wanda-s10-pia
plawanrath/mistral-7b-instruct-v0.3-wanda-s10-pia is a 7 billion parameter instruction-tuned Mistral-7B-Instruct-v0.3 model that has undergone Wanda pruning at a 10% target sparsity. Developed by Plawan Kumar Rath, 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-oriented nature and documented bias amplification.
Loading preview...
Model Overview
This model, plawanrath/mistral-7b-instruct-v0.3-wanda-s10-pia, is a pruned version of the Mistral-7B-Instruct-v0.3 base model, developed by Plawan Kumar Rath. It utilizes the wanda (Activation-aware unstructured pruning) method with a target sparsity of 10%, achieving an actual sparsity of 6.53% by zeroing 455,671,808 parameters. This model is explicitly a research artifact for studying the impact of weight pruning on fairness, as detailed in the paper "Weight Pruning Amplifies Bias: A Multi-Method Study of Compressed LLMs for Edge AI" (IEEE AIIoT 2026).
Key Findings & Characteristics
- Bias Amplification: The primary finding is that Wanda pruning at this sparsity level induces measurable bias amplification on the BBQ benchmark, with a new-bias-emergence rate of 0.83%.
- No Performance Benefits: Despite pruning, the model offers no storage savings (on-disk size is identical to the dense base model) and no latency savings on common hardware like Apple Silicon (MLX) due to unstructured sparsity not being exploited by dense GEMM kernels.
- Smart Pruning Paradox: The research highlights that bias amplification can be invisible to perplexity-based evaluations. For instance, 50% Wanda pruning on Mistral-7B raised perplexity by 3.5% but Stereotype Reliance Score by 83.7%, indicating a 24x disparity.
Important Caveats
- Research Artifact Only: This model is not for production use and should not be deployed in user-facing or decision-making systems due to its documented bias amplification.
- Limited Scope: Pruning was applied to linear layers in transformer blocks (attention projections + MLP), while embeddings, LM head, and layer norms remained untouched.