SciBERT

Training compute

8.9×10¹⁹ FLOP

Parameters

110M

Published

Mar 26, 2019

SciBERT is an AI model developed by Allen Institute for AI (United States), first published in March 2019. It works in the language domain, on tasks such as relation extraction, sentiment classification, text classification and named entity recognition (ner).

Training it took an estimated 8.9×10¹⁹ FLOP of compute (estimation method: hardware). The model has 110,000,000 parameters. It was trained on roughly 3.2B datapoints. Training ran on 4 Google TPU v3 for about 168 hours. The compute alone is estimated at $247 in 2023 dollars.

Access: Open weights (unrestricted). Its weights are openly available. The reference paper has 3,705 citations. Epoch AI rates the confidence of this record as confident.

Full record

Organization: Allen Institute for AI
Country of organization: United States
Domain: Language
Task: Relation extraction, Sentiment classification, Text classification, Named entity recognition (NER)
Training compute: 8.9×10¹⁹ FLOP
Compute estimation method: Hardware
Parameters: 110,000,000
Dataset size: 3.2B
Training hardware: Google TPU v3
Chips used: 4
Training time: 168 h
Chip-hours: 672
Training power draw: 3.7 kW
Training cost (2023 USD): $247
Model accessibility: Open weights (unrestricted)
Open weights: Yes
Citations: 3,705
Epoch confidence: Confident