en.wikipedia.org/wiki/EleutherAI
3 corrections found
3 July 2020 (2020-07-03)
EleutherAI’s own one-year retrospective places the founding of the LibreAI/EleutherAI Discord server on July 7, 2020, not July 3, 2020.
Full reasoning
EleutherAI’s official retrospective was published on July 7, 2021 and says that “on this very day one year ago, the ‘LibreAI’ discord server was founded.” That places the founding date on July 7, 2020. The same post also narrates the July 2 discussion that led to the project and the later July 28 rename to “EleutherAI,” which makes the article’s specific infobox date of July 3, 2020 inconsistent with EleutherAI’s own history.
1 source
- What A Long, Strange Trip It's Been: EleutherAI One Year Retrospective | EleutherAI Blog
July 7, 2021 ... 'on this very day one year ago, the "LibreAI" discord server was founded.' ... Connor 2020-07-28 ... 'we've decided on our new name: EleutherAI'
The numbers of parameters in EleutherAI's GPT-Neo model series are 125 million, 1.3 billion, 2.7 billion, 6 billion, and 20 billion.
This mixes several different EleutherAI model families together. GPT-Neo models included 125M, 350M, 1.3B, and 2.7B; 6B is GPT-J, and 20B is GPT-NeoX-20B.
Full reasoning
EleutherAI’s own gpt-neo repository lists GPT-Neo 125M, 350M, 1.3B, and 2.7B. It does not define 6B or 20B as GPT-Neo variants. Separately, the official GPT-J model card states that “GPT-J” refers to the class of model, while “6B” represents the number of trainable parameters, and the GPT-NeoX repository identifies GPT-NeoX-20B as a distinct pretrained model. So this sentence is incorrect in two ways: it omits the 350M GPT-Neo model and it incorrectly folds GPT-J-6B and GPT-NeoX-20B into the GPT-Neo series.
3 sources
- EleutherAI/gpt-neo README
We're proud to release two pretrained GPT-Neo models ... 1.3B ... 2.7B ... GPT-Neo 125M ... GPT-Neo 350M ... GPT-Neo 1.3B ... GPT-Neo 2.7B.
- EleutherAI GPT-J 6B model card
"GPT-J" refers to the class of model, while "6B" represents the number of trainable parameters.
- EleutherAI/gpt-neox README
Pretrained Models ... GPT-NeoX-20B ... GPT-NeoX-20B is a 20 billion parameter autoregressive language model trained on the Pile.
GPT-J (6B): released in March 2021, it was the largest open-source GPT-3-style language model in the world at the time of release.
GPT-J-6B was not released in March 2021. EleutherAI’s own retrospective records its release on June 9, 2021.
Full reasoning
EleutherAI’s official one-year retrospective gives a dated announcement for GPT-J: “Ben Wang 2021-06-09 ... Today we are releasing the weights for GPT-J-6B.” That directly contradicts the article’s statement that GPT-J was “released in March 2021.”
The retrospective also separately says the GPT-Neo 1.3B and 2.7B models were released on March 21, 2021, which likely explains how the March date was mistakenly carried over to GPT-J.
2 sources
- What A Long, Strange Trip It's Been: EleutherAI One Year Retrospective | EleutherAI Blog
Ben Wang 2021-06-09 '@everyone Today we are releasing the weights for GPT-J-6B...' ... 'we finally got around to releasing them on March 21st, 2021' refers to the 1.3B and 2.7B GPT-Neo models.
- Deploy GPT-J 6B for inference using Hugging Face Transformers and Amazon SageMaker
Published January 11, 2022 ... 'Almost 6 months ago to the day, EleutherAI released GPT-J 6B'.