Aller au contenu
Atlas de l'IA

Phi

Des petits modèles efficaces, bâtis sur des données sélectionnées

Microsoft Modèle Poids ouverts
entréeTexteTexte

Le texte intégral est présenté en anglais ; le titre et le résumé sont localisés.

CE QUE C'EST

Phi is a family of small language models introduced by Microsoft Research in June 2023. Its emphasis is data quality rather than model size: training on filtered “textbook-quality” text lets a few-billion-parameter model approach much larger ones on reasoning and coding. Phi-3 offers 3.8B, 7B and 14B sizes, and the smallest, Phi-3-mini, can run on-device. The models are released as open weights and are often used to probe the capability limits of small models.

Pourquoi il compte

It showed that curating and orchestrating data can substitute for parameter count to a surprising degree: a few-billion-parameter model trained on selected data can approach models an order of magnitude larger on some reasoning tasks.

Caractéristiques clés

Parameters
3.8B / 7B / 14B (Phi-3 family)
Training data
Mainly filtered “textbook-quality” text
Open weights
Yes
Released
2023-06

Capacités

Concepts liés

Produits comparables