Gemma 2B

Training compute

4.5×10²² FLOP

Parameters

2.5B

Published

Feb 21, 2024

Gemma 2B 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 4.5×10²² FLOP of compute (estimation method: operation counting). The model has 2,506,434,560 parameters. It was trained on roughly 3T datapoints. Training ran on 512 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: 4.5×10²² FLOP
Compute estimation method: Operation counting
Parameters: 2,506,434,560
Dataset size: 3T
Training hardware: Google TPU v5e
Chips used: 512
Training power draw: 228.0 kW
Model accessibility: Open weights (restricted use)
Open weights: Yes
Epoch confidence: Confident

Benchmark results

01	Epoch Capabilities Index	ECI Score	93.8
02	PIQA	Score	0.77
03	HellaSwag	Overall accuracy	0.71
04	BoolQ	Score	0.69
05	WinoGrande	Accuracy	0.65
06	TriviaQA	EM	0.53
07	MMLU	EM	0.42
08	ARC AI2	Challenge score	0.42
09	BIG-Bench Hard	Average	0.35
10	GSM8K	EM	0.18