MPT-7B

Training compute

4.2×10²² FLOP

Parameters

7B

Published

May 5, 2023

MPT-7B is an AI model developed by MosaicML (United States), first published in May 2023. It works in the language domain, on tasks such as language modeling/generation and question answering.

Training it took an estimated 4.2×10²² FLOP of compute (estimation method: hardware). The model has 7,000,000,000 parameters. Training ran on 440 NVIDIA A100 SXM4 40 GB for about 228 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 modeling/generation, Question answering
Training compute: 4.2×10²² FLOP
Compute estimation method: Hardware
Parameters: 7,000,000,000
Training hardware: NVIDIA A100 SXM4 40 GB
Chips used: 440
Training time: 228 h
Chip-hours: 100.3K
Training power draw: 350.6 kW
Model accessibility: Open weights (unrestricted)
Open weights: Yes
Epoch confidence: Confident

Benchmark results

01	Epoch Capabilities Index	ECI Score	94.2
02	PIQA	Score	0.81
03	HellaSwag	Overall accuracy	0.76
04	BoolQ	Score	0.75
05	LAMBADA	Score	0.7
06	WinoGrande	Accuracy	0.69
07	TriviaQA	EM	0.62
08	OpenBookQA	Accuracy	0.51
09	ARC AI2	Challenge score	0.43
10	BIG-Bench Hard	Average	0.36