Gemma 7B

Training compute

3.1×10²³ FLOP

Parameters

8.5B

Published

Feb 21, 2024

Gemma 7B is an AI model developed by Google DeepMind (United States), first published in February 2024. It works in the language domain, on tasks such as language modeling/generation, chat, code generation and 2 more.

Training it took an estimated 3.1×10²³ FLOP of compute (estimation method: operation counting,hardware). The model has 8,538,074,112 parameters. It was trained on roughly 6T datapoints. Training ran on 4,096 Google TPU v5e.

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

Full record

Organization: Google DeepMind
Country of organization: United States
Domain: Language
Task: Language modeling/generation, Chat, Code generation, Question answering, Quantitative reasoning
Training compute: 3.1×10²³ FLOP
Compute estimation method: Operation counting, Hardware
Parameters: 8,538,074,112
Dataset size: 6T
Training hardware: Google TPU v5e
Chips used: 4,096
Training power draw: 1.8 MW
Model accessibility: Open weights (restricted use)
Open weights: Yes
Epoch confidence: Confident

Benchmark results

01	Epoch Capabilities Index	ECI Score	111.7
02	BoolQ	Score	0.83
03	HellaSwag	Overall accuracy	0.82
04	PIQA	Score	0.81
05	WinoGrande	Accuracy	0.79
06	OpenBookQA	Accuracy	0.79
07	ARC AI2	Challenge score	0.78
08	TriviaQA	EM	0.72
09	MMLU	EM	0.66
10	BIG-Bench Hard	Average	0.55
11	Adversarial NLI	Score	0.49
12	GSM8K	EM	0.46