brainsteam.co.uk/brainsteam/content/annotations/2023/03/21/1679379947.md

73 lines
3.2 KiB
Markdown
Raw Normal View History

---
date: '2023-03-21T06:25:47'
hypothesis-meta:
created: '2023-03-21T06:25:47.417575+00:00'
document:
title:
- 'GPT-4 and professional benchmarks: the wrong answer to the wrong question'
flagged: false
group: __world__
hidden: false
id: N6BVsMexEe2Z4X92AfjYDg
links:
html: https://hypothes.is/a/N6BVsMexEe2Z4X92AfjYDg
incontext: https://hyp.is/N6BVsMexEe2Z4X92AfjYDg/aisnakeoil.substack.com/p/gpt-4-and-professional-benchmarks
json: https://hypothes.is/api/annotations/N6BVsMexEe2Z4X92AfjYDg
permissions:
admin:
- acct:ravenscroftj@hypothes.is
delete:
- acct:ravenscroftj@hypothes.is
read:
- group:__world__
update:
- acct:ravenscroftj@hypothes.is
tags:
- llm
- openai
- gpt
- ModelEvaluation
target:
- selector:
- endContainer: /div[1]/div[1]/div[2]/div[1]/div[1]/div[1]/article[1]/div[4]/div[1]/div[1]/p[4]/span[2]
endOffset: 300
startContainer: /div[1]/div[1]/div[2]/div[1]/div[1]/div[1]/article[1]/div[4]/div[1]/div[1]/p[4]/span[1]
startOffset: 0
type: RangeSelector
- end: 5998
start: 5517
type: TextPositionSelector
- exact: "To benchmark GPT-4\u2019s coding ability, OpenAI evaluated it on problems\
\ from Codeforces, a website that hosts coding competitions. Surprisingly,\
\ Horace He pointed out that GPT-4 solved 10/10 pre-2021 problems and 0/10\
\ recent problems in the easy category. The training data cutoff for GPT-4\
\ is September 2021. This strongly suggests that the model is able to memorize\
\ solutions from its training set \u2014 or at least partly memorize them,\
\ enough that it can fill in what it can\u2019t recall."
prefix: 'm 1: training data contamination'
suffix: As further evidence for this hyp
type: TextQuoteSelector
source: https://aisnakeoil.substack.com/p/gpt-4-and-professional-benchmarks
text: OpenAI was only able to pass questions available before september 2021 and
failed to answer new questions - strongly suggesting that it has simply memorised
the answers as part of its training
updated: '2023-03-21T06:26:57.441600+00:00'
uri: https://aisnakeoil.substack.com/p/gpt-4-and-professional-benchmarks
user: acct:ravenscroftj@hypothes.is
user_info:
display_name: James Ravenscroft
in-reply-to: https://aisnakeoil.substack.com/p/gpt-4-and-professional-benchmarks
tags:
- llm
- openai
- gpt
- ModelEvaluation
- hypothesis
type: annotation
url: /annotations/2023/03/21/1679379947
---
<blockquote>To benchmark GPT-4s coding ability, OpenAI evaluated it on problems from Codeforces, a website that hosts coding competitions. Surprisingly, Horace He pointed out that GPT-4 solved 10/10 pre-2021 problems and 0/10 recent problems in the easy category. The training data cutoff for GPT-4 is September 2021. This strongly suggests that the model is able to memorize solutions from its training set — or at least partly memorize them, enough that it can fill in what it cant recall.</blockquote>OpenAI was only able to pass questions available before september 2021 and failed to answer new questions - strongly suggesting that it has simply memorised the answers as part of its training