Autoritätsbasiertes Trainingsverfahren für ein Grosssprachenmodell

Details

Veröffentlichungs-Nr.

EP4722943

Anmeldetag

12. Dezember 2024

Veröffentlichung

8. April 2026

Rechtsraum

EP

IPC

G06F16/3329

Abstract

An authority-based training process for a large language model is provided. The process can involve generating corresponding sets of data quality metrics for each sample in a training dataset. The training dataset can encompass a group of topics. The process can also involve generating a corresponding set of authority scores for each sample based on the corresponding sets of data quality metrics. Each authority score can indicate a respective authority level of the sample in relation to a particular topic of the group of topics. The process can further involve training the large language model using a loss function that includes a set of weights. During training, the set of weights can be dynamically adjusted based on the corresponding set of authority scores for each sample in the training dataset. This can produce a large language model that is more accurate than may otherwise be possible.

Anmelder

Firma

Red Hat, LLC

Land

🇺🇸 USA