Diké
Collaborative project

Bias, fairness and ethics of compressed language models

Funded by ANR
(AAPG 2021, 2022-2026)

Rationale

Diké - named after the Greek goddess of justice - is a research project funded by the French National Research Agency (ANR) studying how model compression affects fairness, bias, uncertainty, safety, and model robustness.

Compression methods such as pruning, quantization, and distillation can substantially reduce the computational and memory requirements of language models. However, compressed models are often evaluated mainly through aggregate task performance, leaving open the question of whether model performance on safety benchmarks is preserved.

The Diké project examines how compression affects performance across texts with different demographic groups mentions, model confidence and calibration, safety, multilingual performance, and model robustness, including cases where standard performance metrics remain largely unchanged.
3
Research partners
Hubert Curien Lab · ERIC · Naver Labs Europe
15+
Publications
EMNLP · NAACL · TACL · JMLR · ICASSP · LREC
6
Research internships
Student research across partner institutions
3
Compression techniques
Pruning · Quantization · Distillation
2
Datasets & benchmarks
Toxicity evaluation · Moral alignment
2
Workshops organized
Diké Workshop 2022 · EALM @ TALN 2025
2
PhD theses
Doctoral research within the Diké project
1
Open-source leaderboard
Open Diké Leaderboard

Research objectives

  • Objective 1. To create bilingual datasets in English and French for bias, fairness and ethics evaluation in language models.
  • Objective 2. To devise evaluation metrics and to perform evaluation campaigns on compression techniques.
  • Objective 3. To propose new neural architectures for less biased, fairer and more ethical compression techniques.

Project contributions

Datasets & benchmarks

Introduced new evaluation resources for French language models, including FrenchToxicityPrompts for toxicity evaluation and HistoiresMorales for evaluating moral alignment.

Compression & evaluation

The project studies the impact of pruning, quantization, and distillation on properties beyond aggregate predictive performance, including group fairness, uncertainty, confidence, calibration, multilingual performance, and robustness.

Fair representations

The project develops methods for learning fair and transferable representations, including Wasserstein-based independence objectives designed to reduce dependence on sensitive attributes and transfer fairness across domains.

Fairness-aware compression

The project develops compression methods that explicitly account for fairness, including Fair-GPTQ for fairness-aware quantization and Debias-SparseGPT for bias-aware pruning.

Context

Large-scale models like BERT and GPT-3, while easily accessible, pose scalability challenges due to their immense size—ranging from 110 million to 175 billion parameters. Recent compression techniques aim to improve their usability by employing such methods as:

  • Pruning, a technique that aims at removing weights which are identified as less relevant for a trained model.
  • Weights Quantization, which is compressing each parameter of a model in a lower number of bits.
  • Distillation, consisting in training a smaller model, or “student model” that aims to mimic the predictions of a larger pre-trained model, or “teacher model” to ease training.

In that scientific context, the Diké project builds upon the following key observation. All compression techniques only focus on preserving the model accuracy on a given task. But there is no such thing as a free lunch: if we remove most of these weights (sometimes 99%!) and the accuracy of the model remains the same if not higher, we may be trading for model bias, fairness or ethics without knowing it. We study the hypothesis that what is discarded may be related to harmful side effects.