All-attention network + adaptive span

Training compute

1.2×10¹⁹ FLOP

Parameters

133M

Published

Jul 2, 2019

All-attention network + adaptive span is an AI model developed by Facebook AI Research (United States and France), first published in July 2019. It works in the language domain, on tasks such as language modeling/generation.

Training it took an estimated 1.2×10¹⁹ FLOP of compute (estimation method: hardware,operation counting). The model has 133,000,000 parameters. It was trained on roughly 103M datapoints. Training ran on 64 NVIDIA V100 for about 24 hours.

Access: Unreleased. Its weights are not openly released. The reference paper has 156 citations. Epoch AI rates the confidence of this record as likely.

Full record

Organization: Facebook AI Research
Country of organization: United States, France
Domain: Language
Task: Language modeling/generation
Training compute: 1.2×10¹⁹ FLOP
Compute estimation method: Hardware, Operation counting
Parameters: 133,000,000
Dataset size: 103M
Training hardware: NVIDIA V100
Chips used: 64
Training time: 24 h
Training power draw: 39.5 kW
Model accessibility: Unreleased
Open weights: No
Citations: 156
Epoch confidence: Likely