LLaMA-65B

Training compute

5.5×10²³ FLOP

Parameters

65.2B

Published

Feb 24, 2023

LLaMA-65B is an AI model developed by Meta AI (United States), first published in February 2023. It works in the language domain, on tasks such as language modeling and code generation.

Training it took an estimated 5.5×10²³ FLOP of compute (estimation method: operation counting). The model has 65,200,000,000 parameters. It was trained on roughly 1.4T datapoints. Training ran on 2,048 NVIDIA A100 for about 500 hours. The compute alone is estimated at $578K in 2023 dollars.

Access: Open weights (non-commercial). Its weights are openly available. The reference paper has 19,926 citations. Epoch AI rates the confidence of this record as confident.

Full record

Organization: Meta AI
Country of organization: United States
Domain: Language
Task: Language modeling, Code generation
Training compute: 5.5×10²³ FLOP
Compute estimation method: Operation counting
Parameters: 65,200,000,000
Dataset size: 1.4T
Training hardware: NVIDIA A100
Chips used: 2,048
Training time: 500 h
Chip-hours: 1M
Training power draw: 1.6 MW
Training cost (2023 USD): $578K
Model accessibility: Open weights (non-commercial)
Open weights: Yes
Citations: 19,926
Epoch confidence: Confident

Benchmark results

01	Epoch Capabilities Index	ECI Score	109.9
02	BoolQ	Score	0.87
03	TriviaQA	EM	0.86
04	HellaSwag	Overall accuracy	0.84
05	PIQA	Score	0.83
06	LAMBADA	Score	0.78
07	WinoGrande	Accuracy	0.77
08	ARC AI2	Challenge score	0.69
09	MMLU	EM	0.63
10	OpenBookQA	Accuracy	0.6
11	BIG-Bench Hard	Average	0.58
12	GSM8K	EM	0.54