Towards an Ethical Compression of Large Language Models
Résumé
This proposal explores the fairness of compressed large language models (LLMs). We focus on the ethical implications of applying efficient compression techniques, particularly quantization, to generative LLMs, motivated by recent studies. While quantization enhances inference efficiency, marked by existing works, we primarily focus on understanding its effects on token-level confidence and predictive probability distributions in our research. We also identify significant influences on LLM behaviour during text generation, shedding light on potential biases and ethical concerns. We have determined the difference in output probability distributions after compression and aim to use this observation to propose a debiasing quantization approach.
Origine | Fichiers produits par l'(les) auteur(s) |
---|