Phi-1.5

Training compute

1.2×10²¹ FLOP

Parameters

1.3B

Published

Sep 11, 2023

Phi-1.5 is an AI model developed by Microsoft (United States), first published in September 2023. It works in the language domain, on tasks such as language generation.

Training it took an estimated 1.2×10²¹ FLOP of compute (estimation method: operation counting,hardware). The model has 1,300,000,000 parameters. It was trained on roughly 30B datapoints. Training ran on NVIDIA A100 SXM4 40 GB for about 192 hours.

Access: Open weights (unrestricted). Its weights are openly available. The reference paper has 647 citations. Epoch AI rates the confidence of this record as confident.

Full record

Organization: Microsoft
Country of organization: United States
Domain: Language
Task: Language generation
Training compute: 1.2×10²¹ FLOP
Compute estimation method: Operation counting, Hardware
Parameters: 1,300,000,000
Dataset size: 30B
Training hardware: NVIDIA A100 SXM4 40 GB
Training time: 192 h
Model accessibility: Open weights (unrestricted)
Open weights: Yes
Citations: 647
Epoch confidence: Confident

Benchmark results

01	Epoch Capabilities Index	ECI Score	91.2
02	BoolQ	Score	0.76
03	WinoGrande	Accuracy	0.73
04	HellaSwag	Overall accuracy	0.48
05	ARC AI2	Challenge score	0.44
06	MMLU	EM	0.38
07	OpenBookQA	Accuracy	0.37