Joe Ornstein
@joeornstein.bsky.social
Political Science @ UGA
https://joeornstein.github.io/
https://joeornstein.github.io/
3. 🦜R package alert! 🦜
With promptr, R users can easily format and complete few-shot LLM prompts for document labeling and scaling tasks.
cran.r-project.org/web/packages...
With promptr, R users can easily format and complete few-shot LLM prompts for document labeling and scaling tasks.
cran.r-project.org/web/packages...
promptr: Format and Complete Few-Shot LLM Prompts
Format and submit few-shot prompts to OpenAI's Large Language Models (LLMs). Designed to be particularly useful for text classification problems in the social sciences. Methods are described in Ornste...
cran.r-project.org
January 14, 2025 at 7:39 PM
3. 🦜R package alert! 🦜
With promptr, R users can easily format and complete few-shot LLM prompts for document labeling and scaling tasks.
cran.r-project.org/web/packages...
With promptr, R users can easily format and complete few-shot LLM prompts for document labeling and scaling tasks.
cran.r-project.org/web/packages...
2a. If you're a social scientist using crowd-coding platforms to label documents in 2025, you're spending 1,000 times more money to ask someone else to put your text into an LLM for you.
January 14, 2025 at 7:39 PM
2a. If you're a social scientist using crowd-coding platforms to label documents in 2025, you're spending 1,000 times more money to ask someone else to put your text into an LLM for you.
2. When we started the project in fall 2021, our analyses cost a few hundred dollars in API fees. Today, the the same tasks would cost around $3. That's about 1,200 times cheaper than performing the same tasks on crowd-coding platforms.
January 14, 2025 at 7:39 PM
2. When we started the project in fall 2021, our analyses cost a few hundred dollars in API fees. Today, the the same tasks would cost around $3. That's about 1,200 times cheaper than performing the same tasks on crowd-coding platforms.
1. It's remarkable that, in 2025, GPT-3 still performs as well if not better than GPT-4 and its offshoots at our document labeling and scaling tasks. RLHF is great for making chatbots, but for text-as-data tasks you're often better off with the base models.
January 14, 2025 at 7:39 PM
1. It's remarkable that, in 2025, GPT-3 still performs as well if not better than GPT-4 and its offshoots at our document labeling and scaling tasks. RLHF is great for making chatbots, but for text-as-data tasks you're often better off with the base models.