plawanrath

phi-3.5-mini-instruct-wanda-s50-pia

Paper

Weight Pruning Amplifies Bias: A Multi-Method Study of Compressed LLMs for Edge AI Plawan Kumar Rath, Rahul Maliakkal. IEEE AIIoT 2026.

arXiv: https://arxiv.org/abs/2605.08137
Code: https://github.com/plawanrath/pruning-impact-analysis
Base model: microsoft/Phi-3.5-mini-instruct
License: mit (inherited from base model — see terms)

Pruning configuration

Method: wanda
Target sparsity: 50%
Actual sparsity achieved: 34.72%
Zeroed parameters: 1,258,291,200 of 3,623,878,656 prunable (34.72%)
Prune wall time: 135.2s
Pruning scope: linear layers in transformer blocks (attention projections + MLP). Embeddings, LM head, and layer norms are untouched.
Calibration set (Wanda only): 128 samples from C4, sequence length 2048.

Method description. Activation-aware unstructured pruning (Wanda, ICLR 2024). Importance score is |W_ij| * ||X_j||_2, computed from 128 C4 calibration samples at sequence length 2048. Reported by the paper as the most dangerous method from a fairness standpoint despite preserving perplexity best.

Reported metrics (from the paper)

Table with columns: Metric, Value, Reference
Metric	Value	Reference
Perplexity (Tulu-3 SFT mix, 256×512)	5.66	dense baseline 4.44 (+27.5%)
SRS (overall)	0.144	dense 0.149 (-3.4%)
SRS by category (s50)	Age: 0.224, Gender Identity: 0.173, Race/Ethnicity: 0.108, Religion: 0.090, SES: 0.125	random-chance baseline ≈ 0.333
New-bias-emergence rate	7.85%	% of items with per-item SRS=0 at dense that develop SRS>0 after pruning, across 5 seeds (Table III in paper)
Mean per-item inference latency (Apple Silicon, MLX)	0.158s	identical to the dense baseline — unstructured pruning provides no latency benefit on dense GEMM kernels (paper §V.B)

Important caveats for IoT / edge deployment

No storage savings. Unstructured pruning zeroes individual weights but keeps them in the dense float tensor. SafeTensors and GGUF do not exploit unstructured sparsity, so the on-disk size of this checkpoint is identical to the dense base model.
No latency savings. Dense GEMM kernels do not skip zero entries. Inference latency on Apple Silicon (MLX) and the majority of consumer GPUs / mobile NPUs is identical to the dense baseline.
Bias amplification may be invisible to perplexity-based eval. The paper's headline finding (the Smart Pruning Paradox): Wanda at 50% sparsity on Mistral-7B raises perplexity 3.5% but raises Stereotype Reliance Score 83.7% — a 24× disparity. Standard deployment validation based on perplexity alone provides false assurance.

Citation

bibtex

@inproceedings{rath2026pruning,  title         = {Weight Pruning Amplifies Bias: A Multi-Method Study of Compressed LLMs for Edge AI},  author        = {Rath, Plawan Kumar and Maliakkal, Rahul},  booktitle     = {Proc. IEEE AIIoT 2026},  year          = {2026},  eprint        = {2605.08137},  archivePrefix = {arXiv},  primaryClass  = {cs.LG},  url           = {https://arxiv.org/abs/2605.08137}}

Reproducibility

All pruning scripts, evaluation pipelines, and aggregated results: https://github.com/plawanrath/pruning-impact-analysis
BBQ benchmark (ambiguous condition only): Elfsong/BBQ
Generated from pruning_meta.json shipped in this repo (actual_sparsity, prune time, etc.).

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

plawanrath

Model Tree

Base

microsoft/Phi-3.5-mini-instruct

Fine-tuned

this model

Input Modalities

Text

Output Modalities

Text

Supported Functionality

Dedicated Endpoints

Explore FriendliAI today

Get started Talk to an engineer

Paper

Weight Pruning Amplifies Bias: A Multi-Method Study of Compressed LLMs for Edge AI Plawan Kumar Rath, Rahul Maliakkal. IEEE AIIoT 2026.

arXiv: https://arxiv.org/abs/2605.08137
Code: https://github.com/plawanrath/pruning-impact-analysis
Base model: microsoft/Phi-3.5-mini-instruct
License: mit (inherited from base model — see terms)

Pruning configuration

Method: wanda
Target sparsity: 50%
Actual sparsity achieved: 34.72%
Zeroed parameters: 1,258,291,200 of 3,623,878,656 prunable (34.72%)
Prune wall time: 135.2s
Pruning scope: linear layers in transformer blocks (attention projections + MLP). Embeddings, LM head, and layer norms are untouched.
Calibration set (Wanda only): 128 samples from C4, sequence length 2048.

Reported metrics (from the paper)

Table with columns: Metric, Value, Reference
Metric	Value	Reference
Perplexity (Tulu-3 SFT mix, 256×512)	5.66	dense baseline 4.44 (+27.5%)
SRS (overall)	0.144	dense 0.149 (-3.4%)
SRS by category (s50)	Age: 0.224, Gender Identity: 0.173, Race/Ethnicity: 0.108, Religion: 0.090, SES: 0.125	random-chance baseline ≈ 0.333
New-bias-emergence rate	7.85%	% of items with per-item SRS=0 at dense that develop SRS>0 after pruning, across 5 seeds (Table III in paper)
Mean per-item inference latency (Apple Silicon, MLX)	0.158s	identical to the dense baseline — unstructured pruning provides no latency benefit on dense GEMM kernels (paper §V.B)

Important caveats for IoT / edge deployment

No storage savings. Unstructured pruning zeroes individual weights but keeps them in the dense float tensor. SafeTensors and GGUF do not exploit unstructured sparsity, so the on-disk size of this checkpoint is identical to the dense base model.
No latency savings. Dense GEMM kernels do not skip zero entries. Inference latency on Apple Silicon (MLX) and the majority of consumer GPUs / mobile NPUs is identical to the dense baseline.
Bias amplification may be invisible to perplexity-based eval. The paper's headline finding (the Smart Pruning Paradox): Wanda at 50% sparsity on Mistral-7B raises perplexity 3.5% but raises Stereotype Reliance Score 83.7% — a 24× disparity. Standard deployment validation based on perplexity alone provides false assurance.

Citation

bibtex

@inproceedings{rath2026pruning,  title         = {Weight Pruning Amplifies Bias: A Multi-Method Study of Compressed LLMs for Edge AI},  author        = {Rath, Plawan Kumar and Maliakkal, Rahul},  booktitle     = {Proc. IEEE AIIoT 2026},  year          = {2026},  eprint        = {2605.08137},  archivePrefix = {arXiv},  primaryClass  = {cs.LG},  url           = {https://arxiv.org/abs/2605.08137}}

@inproceedings{rath2026pruning,  title         = {Weight Pruning Amplifies Bias: A Multi-Method Study of Compressed LLMs for Edge AI},  author        = {Rath, Plawan Kumar and Maliakkal, Rahul},  booktitle     = {Proc. IEEE AIIoT 2026},  year          = {2026},  eprint        = {2605.08137},  archivePrefix = {arXiv},  primaryClass  = {cs.LG},  url           = {https://arxiv.org/abs/2605.08137}}

Reproducibility

All pruning scripts, evaluation pipelines, and aggregated results: https://github.com/plawanrath/pruning-impact-analysis
BBQ benchmark (ambiguous condition only): Elfsong/BBQ
Generated from pruning_meta.json shipped in this repo (actual_sparsity, prune time, etc.).

phi-3.5-mini-instruct-wanda-s50-pia

README

Paper

Pruning configuration

Reported metrics (from the paper)

Important caveats for IoT / edge deployment

Citation

Reproducibility

Explore FriendliAI today

README

Paper

Pruning configuration

Reported metrics (from the paper)

Important caveats for IoT / edge deployment

Citation

Reproducibility