Llama 2-34B

Training compute

4.1×10²³ FLOP

Parameters

34B

Published

Jul 18, 2023

Llama 2-34B is an AI model developed by Meta AI (United States), first published in July 2023. It works in the language domain, on tasks such as language modeling.

Training it took an estimated 4.1×10²³ FLOP of compute (estimation method: hardware,operation counting). The model has 34,000,000,000 parameters. It was trained on roughly 2T datapoints. Training ran on NVIDIA A100 SXM4 80 GB. The compute alone is estimated at $600K in 2023 dollars.

Access: Unreleased. Its weights are not openly released. The reference paper has 16,911 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
Training compute: 4.1×10²³ FLOP
Compute estimation method: Hardware, Operation counting
Parameters: 34,000,000,000
Dataset size: 2T
Training hardware: NVIDIA A100 SXM4 80 GB
Training cost (2023 USD): $600K
Model accessibility: Unreleased
Open weights: No
Citations: 16,911
Epoch confidence: Confident

Benchmark results

01	Epoch Capabilities Index	ECI Score	104.9
02	TriviaQA	EM	0.85
03	BoolQ	Score	0.84
04	PIQA	Score	0.82
05	WinoGrande	Accuracy	0.77
06	MMLU	EM	0.63
07	OpenBookQA	Accuracy	0.58
08	ARC AI2	Challenge score	0.55
09	BIG-Bench Hard	Average	0.44
10	GSM8K	EM	0.42