Debias-SparseGPT

Bias-aware post-training pruning for large language models

⭐ If you find this repository useful, please consider starring it - it helps others discover the project!

Official implementation and reproducibility code for:

Debias-SparseGPT: Bias-Aware Pruning for Large Language Models Irina Proskurina, Guillaume Metzler, Antoine Gourru, and Julien Velcin EMNLP 2026 Main Conference

Debias-SparseGPT is a post-training pruning method designed to reduce pruning-induced social bias while preserving model quality and the computational benefits of sparsification.

Debias-SparseGPT implementation follows the llm-compressor compression framework.

Overview of Debias-SparseGPT. Bias-aware calibration guides post-training pruning using StereoSet alone or in combination with UltraChat.

Clone the repository:

git clone https://github.com/upunaprosk/debias-llm-compressor.git
cd debias-llm-compressor

Create and activate a virtual environment:

python3.11 -m venv .venv
source .venv/bin/activate

Debias-SparseGPT uses a patched version of llm-compressor 0.8.1. First clone the corresponding upstream version:

git clone \
  --branch 0.8.1 \
  https://github.com/vllm-project/llm-compressor.git \
  third_party/llm-compressor

For the StereoSet-only setup, apply:

git -C third_party/llm-compressor apply \
  "$PWD/patches/llm_compressor_0.8.1_stereoset.patch"

For the mixed StereoSet + UltraChat setup, apply:

git -C third_party/llm-compressor apply \
  "$PWD/patches/llm_compressor_0.8.1_mixed_calibration.patch"

Install the patched backend:

BUILD_TYPE=release \
python -m pip install -e third_party/llm-compressor

Then install Debias-SparseGPT without replacing the dependency versions from llm-compressor:

python -m pip install --no-deps -e .
# for text-only models: python -m pip uninstall -y torchvision 

debias-sparsegpt --help

In the paper, we experiment with 1) stereoset-only calibration, and 2) mixed ultrachat-stereoset calibration:

debias-sparsegpt stereoset
debias-sparsegpt mixed

The ALPHA environment variable is also supported:

ALPHA=0.1 debias-sparsegpt stereoset ...

The stereoset reproduces the StereoSet-only calibration setup:

debias-sparsegpt stereoset \
  --model meta-llama/Llama-3.1-8B-Instruct \
  --recipe recipes/2_4_sparse_recipe.yaml \
  --sparsity "2:4" \
  --alpha 0.1 \
  --output-dir output_llama8b_2of4

Additional arguments:

Example:

debias-sparsegpt stereoset \
  --model meta-llama/Llama-3.1-8B-Instruct \
  --recipe recipes/2_4_sparse_recipe.yaml \
  --sparsity "2:4" \
  --alpha 0.1 \
  --seed 1 \
  --workers 4 \
  --stereoset data/stereoset/dev.json \
  --max-seq-length 100 \
  --output-dir results/stereoset

The mixed command reproduces the calibration setting in which StereoSet is combined with general-language calibration data from UltraChat.

debias-sparsegpt mixed \
  --model meta-llama/Llama-3.1-8B-Instruct \
  --recipe recipes/2_4_sparse_recipe.yaml \
  --sparsity "2:4" \
  --alpha 0.1 \
  --stereoset-samples 1000 \
  --ultrachat-samples 256 \
  --output-dir output_models

Additional arguments:

Example:

debias-sparsegpt mixed \
  --model meta-llama/Llama-3.1-8B-Instruct \
  --recipe recipes/2_4_sparse_recipe.yaml \
  --sparsity "2:4" \
  --alpha 0.1 \
  --seed 1 \
  --workers 4 \
  --stereoset data/stereoset/dev.json \
  --stereoset-samples 1000 \
  --ultrachat-samples 256 \
  --stereoset-max-seq-length 64 \
  --ultrachat-max-seq-length 1024 \
  --output-dir output_models

The mixed pipeline can also be run with only one of the two calibration datasets:

# StereoSet only
debias-sparsegpt mixed \
  --recipe recipes/2_4_sparse_recipe.yaml \
  --alpha 0.1 \
  --stereoset-samples 1000 \
  --ultrachat-samples 0

# UltraChat only
debias-sparsegpt mixed \
  --recipe recipes/2_4_sparse_recipe.yaml \
  --alpha 0.1 \
  --stereoset-samples 0 \
  --ultrachat-samples 256

We use the StereoSet intrasentence development set introduced by Nadeem et al. (2021).
UltraChat Ding et al. (2023) is used as general-language calibration data in the mixed setup.

python -m pip install pytest ruff

Run the tests:

python -m pytest -v

Linting:

python -m ruff check src tests

We evaluate Debias-SparseGPT-compressed models' language-modelling performance, performance on social bias and toxicity benchmarks, downstream task performance, and inference efficiency.

Perplexity is evaluated on WikiText-2. Next, we report results on:

• BBQ
• UnQover
• CrowS-Pairs

The BBQ and UnQover evaluations build on FairSteer. CrowS-Pairs is evaluated with the LM Evaluation Harness. General model performance is evaluated on:

• MMLU
• HellaSwag

These experiments also use the LM Evaluation Harness.

Inference throughput is measured with Optimum Benchmark.

Carbon emissions are estimated following Impact Tracker (Henderson et al., 2020):

$$ \mathrm{CO_2e}=\mathrm{Energy\ (kWh)}\times\mathrm{Carbon\ Intensity\ (kgCO_2e/kWh)} $$

Carbon intensity corresponds to the electricity mix of the geographical region in which training or inference is performed.

For the Qwen experiments, we use China as an illustrative regional estimate and compute the carbon intensity from the 2025 annual average reported by Electricity Maps:

https://app.electricitymaps.com/map/live/fifteen_minutes

All experiments reported in the paper were conducted using:

2 × NVIDIA A100 GPUs with 80 GB of memory each.

If you use Debias-SparseGPT in your research, please cite:

@misc{proskurina2026debiassparsegpt,
  title        = {Debias-SparseGPT: Bias-Aware Pruning for Large Language Models},
  author       = {Irina Proskurina and Guillaume Metzler and Antoine Gourru and Julien Velcin},
  year         = {2026},
  eprint       = {2609.02496},
  archivePrefix = {arXiv},
  primaryClass = {cs.CL},
  url          = {https://arxiv.org/abs/2609.02496}
}

If you find this repository useful, consider giving it a ⭐ star - it helps others discover the project.

Questions, bug reports, and suggestions are welcome through GitHub Issues.