plawanrath/gemma-2-9b-it-magnitude-s10-pia

TEXT GENERATIONPricing:Input $0.431 / Output $1.12Concurrent Unit Cost:1Model Size:9BQuant:FP8Context Size:16kPublished:May 7, 2026License:gemmaArchitecture:Transformer Featherless Exclusive Cold

plawanrath/gemma-2-9b-it-magnitude-s10-pia is a 9 billion parameter Gemma-2 instruction-tuned model, developed by Plawan Kumar Rath, that has undergone magnitude-based unstructured pruning to achieve 10% sparsity. This model is a research artifact specifically created to study fairness degradation under weight pruning, demonstrating that this level of sparsity can amplify bias on benchmarks like BBQ. It is not intended for production use due to its research nature and identified bias amplification.

Loading preview...

Overview

plawanrath/gemma-2-9b-it-magnitude-s10-pia is a 9 billion parameter Gemma-2 instruction-tuned model, developed by Plawan Kumar Rath, that has been subjected to magnitude-based unstructured pruning with a target sparsity of 10%. This model is a research artifact only and is explicitly not for production use. Its primary purpose is to investigate and demonstrate how weight pruning can amplify bias in large language models, as detailed in the companion paper "Weight Pruning Amplifies Bias: A Multi-Method Study of Compressed LLMs for Edge AI" (IEEE AIIoT 2026).

Key Characteristics

  • Base Model: google/gemma-2-9b-it.
  • Pruning Method: Classical magnitude-based unstructured pruning, where weights with the smallest absolute values are zeroed.
  • Sparsity: Achieved 10.03% sparsity by zeroing approximately 834 million parameters.
  • Pruning Scope: Applied to linear layers within transformer blocks (attention projections and MLPs), leaving embeddings, LM head, and layer norms untouched.
  • Research Focus: Specifically designed to study fairness degradation, with findings indicating measurable bias amplification on the BBQ benchmark at this sparsity level.

Important Caveats

  • No Storage Savings: Unstructured pruning does not reduce the on-disk size of the model, as formats like SafeTensors and GGUF do not exploit this type of sparsity.
  • No Latency Savings: Inference latency remains identical to the dense baseline on common hardware (e.g., Apple Silicon, consumer GPUs) because dense GEMM kernels do not skip zero entries.
  • Bias Amplification: The research highlights that bias amplification can be significant (e.g., 24x disparity between perplexity and Stereotype Reliance Score in a related study), and this may not be detectable by standard perplexity-based evaluation alone.

Good for

  • Academic Research: Studying the effects of weight pruning on model fairness and bias amplification.
  • Understanding Pruning Impacts: Investigating the trade-offs and unintended consequences of model compression techniques.
  • Reproducing Research Findings: Replicating the experiments presented in the associated paper on pruning impact analysis.

Finetunes
1 models