gradiend 0.1.0

GRADIEND: Gradient-based Targeted Feature Learning within Neural Networks

GRADIEND

Gradient-based targeted feature learning within neural networks. Learn where a language model encodes a feature (e.g. gender, race) and rewrite the model to strengthen or weaken it—for example debias it—while keeping other behaviour unchanged.

Paper: GRADIEND: Feature Learning within Neural Networks Exemplified through Biases

---

Install

pip install gradiend

Optional extras (install one or combine with e.g. gradiend[data,plot]):

Tag	Command	Includes
recommended	pip install gradiend[recommended]	full data generation, training, and plotting
data	pip install gradiend[data]	for advanced data generation and import
plot	pip install gradiend[plot]	for plotting support
dev	pip install gradiend[dev]	for building docs and running tests

---

Links

Documentation	aieng-lab.github.io/gradiend
Source & issues	github.com/aieng-lab/gradiend
Example scripts & notebooks	gradiend/examples — start_workflow, english_pronouns, gender_de, gender_en, race_religion, etc.
Datasets (Hugging Face)	de-gender-case-articles, gradiend_race_data, gradiend_religion_data, biasneutral, geneutral
Pre-trained GRADIEND models	bert-base-cased-gradiend-gender-debiased, gpt2-gradiend-gender-debiased, Llama-3.2-3B-gradiend-gender-debiased

---

Citation

@misc{drechsel2025gradiend,
  title={{GRADIEND}: Feature Learning within Neural Networks Exemplified through Biases},
  author={Jonathan Drechsel and Steffen Herbold},
  year={2025},
  eprint={2502.01406},
  archivePrefix={arXiv},
  primaryClass={cs.LG},
  url={https://arxiv.org/abs/2502.01406},
}

German definite articles study:

@misc{drechsel2026understanding,
  title={Understanding or Memorizing? A Case Study of German Definite Articles in Language Models},
  author={Jonathan Drechsel and Erisa Bytyqi and Steffen Herbold},
  year={2026},
  eprint={2601.09313},
  archivePrefix={arXiv},
  primaryClass={cs.CL},
  url={https://arxiv.org/abs/2601.09313},
}

---

License

Apache 2.0. See LICENSE on GitHub.