Retired Jan 1, 2023. Superseded by GPT-3 and later models. Still available on HuggingFace.“Too dangerous to release. Now a punchline.”
GPT-2 performance data on Rival is based on blind head-to-head community voting. Overall win rate: 9.1% across 33 duels. All vote data is part of Rival's open dataset of 21,000+ human preference judgments across 200+ AI models. Model responses are curated from 7 challenges.
GPT-2 is good. We've said that. We stand by it. But we'd be doing you a disservice if we didn't show you these.
Retired Jan 1, 2023. Superseded by GPT-3 and later models. Still available on HuggingFace.“Too dangerous to release. Now a punchline.”
GPT-2 performance data on Rival is based on blind head-to-head community voting. Overall win rate: 9.1% across 33 duels. All vote data is part of Rival's open dataset of 21,000+ human preference judgments across 200+ AI models. Model responses are curated from 7 challenges.
GPT-2 is good. We've said that. We stand by it. But we'd be doing you a disservice if we didn't show you these.
A direct scale-up of GPT-1 with 1.5 billion parameters, trained on 8 million web pages. Known for its ability to generate coherent text, sometimes indistinguishable from humans, but could be repetitive.
Taste is judged on an uncapped scale where 100 is the reference, originality first. The space past 100 is the craft today's models rarely reach.
Unique words vs. total words. Higher = richer vocabulary.
Average words per sentence.
"Might", "perhaps", "arguably" per 100 words.
**Bold** markers per 1,000 characters.
Bullet and numbered list items per 1,000 characters.
Markdown headings per 1,000 characters.
Emoji per 1,000 characters.
"However", "moreover", "furthermore" per 100 words.
7 outputs from GPT-2
A direct scale-up of GPT-1 with 1.5 billion parameters, trained on 8 million web pages. Known for its ability to generate coherent text, sometimes indistinguishable from humans, but could be repetitive.
Taste is judged on an uncapped scale where 100 is the reference, originality first. The space past 100 is the craft today's models rarely reach.
Unique words vs. total words. Higher = richer vocabulary.
Average words per sentence.
"Might", "perhaps", "arguably" per 100 words.
**Bold** markers per 1,000 characters.
Bullet and numbered list items per 1,000 characters.
Markdown headings per 1,000 characters.
Emoji per 1,000 characters.
"However", "moreover", "furthermore" per 100 words.
7 outputs from GPT-2