MPT-30B

Training compute

1.9×10²³ FLOP

Parameters

30B

Published

Jun 22, 2023

MPT-30B is an AI model developed by MosaicML (United States), first published in June 2023. It works in the language domain, on tasks such as language generation and code generation.

Training it took an estimated 1.9×10²³ FLOP of compute (estimation method: operation counting). The model has 30,000,000,000 parameters. It was trained on roughly 1.1T datapoints. Training ran on 512 NVIDIA H100 SXM5 80GB for about 278 hours.

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

Full record

Organization: MosaicML
Country of organization: United States
Domain: Language
Task: Language generation, Code generation
Training compute: 1.9×10²³ FLOP
Compute estimation method: Operation counting
Parameters: 30,000,000,000
Dataset size: 1.1T
Training hardware: NVIDIA H100 SXM5 80GB
Chips used: 512
Training time: 278 h
Chip-hours: 142.5K
Training power draw: 713.2 kW
Model accessibility: Open weights (unrestricted)
Open weights: Yes
Epoch confidence: Confident

Benchmark results

01	Epoch Capabilities Index	ECI Score	100.3
02	PIQA	Score	0.82
03	BoolQ	Score	0.79
04	TriviaQA	EM	0.74
05	WinoGrande	Accuracy	0.71
06	OpenBookQA	Accuracy	0.52
07	ARC AI2	Challenge score	0.51
08	MMLU	EM	0.48
09	BIG-Bench Hard	Average	0.38
10	GSM8K	EM	0.16