Phi-2

Training compute

2.3×10²² FLOP

Parameters

2.7B

Published

Dec 12, 2023

Phi-2 is an AI model developed by Microsoft (United States), first published in December 2023. It works in the language domain, on tasks such as language generation and code generation.

Training it took an estimated 2.3×10²² FLOP of compute (estimation method: operation counting,hardware). The model has 2,700,000,000 parameters. It was trained on roughly 250B datapoints. Training ran on NVIDIA A100 for about 336 hours.

Access: Open weights (unrestricted). Its weights are openly available. Epoch AI rates the confidence of this record as confident.

Full record

Organization: Microsoft
Country of organization: United States
Domain: Language
Task: Language generation, Code generation
Training compute: 2.3×10²² FLOP
Compute estimation method: Operation counting, Hardware
Parameters: 2,700,000,000
Dataset size: 250B
Training hardware: NVIDIA A100
Training time: 336 h
Model accessibility: Open weights (unrestricted)
Open weights: Yes
Epoch confidence: Confident

Benchmark results

01	Epoch Capabilities Index	ECI Score	107.6
02	ARC AI2	Challenge score	0.76
03	OpenBookQA	Accuracy	0.74
04	BIG-Bench Hard	Average	0.59
05	MMLU	EM	0.58
06	WinoGrande	Accuracy	0.55
07	HellaSwag	Overall accuracy	0.54
08	TriviaQA	EM	0.45
09	Adversarial NLI	Score	0.42