Building on @sewon__min et al.'s find that multiple-choice examples with random answers barely harm performance, @BoshiWang2 et al. find you can achieve 80-90% the gains of chain-of-thought using only examples with invalid reasoning.
LLMS
-
LangChain v0.0.43: Weaviate and Pinecone vectorstore support
By
–
v0.0.43 @weaviate_io vectorstore support! h/t @CShorten30 for help here – keep out for more! @pinecone vectorstore support – @cameroncohen Improvements to text splitter And better docs for data augmented generation in general! https://
langchain.readthedocs.io/en/latest/exam
ples/data_augmented_generation.html
… -
Anthropic Shares RLHF Datasets Publicly on Hugging Face
By
–
Very cool and grateful to see @AnthropicAI sharing datasets publicly, especially around RLHF! https://
huggingface.co/Anthropic -
ChatGPT: Everything You Really Need To Know Simply
By
–
ChatGPT: Everything You Really Need To Know (In Simple Terms) https://
forbes.com/sites/bernardm
arr/2022/12/21/chatgpt-everything-you-really-need-to-know-in-simple-terms/
… #AI #MachineLearning #DataScience #IoT #100DaysofCode #womenwhocode #serverless #Robots @FmFrancoise @CatherineAdenle @Shi4Tech @enilev #BigData #Analytics #DeepLearning #DigitalTransformation -
Extending arithmetic findings to unseen pre-training tasks
By
–
Cool work! In addition to arithmetic, I wonder how much these findings apply to a task that is not seen much in pre-training, such as last letter concatenation or coin flip
-
Large language models quantifiably tell us what we want to hear.
By
–
Large language models are, quantifiably, telling us what we want to hear.
-
ChatGPT Basics: Everything You Need to Know Simply
By
–
ChatGPT: Everything You Really Need to Know (In Simple Terms) Get the basics of #ChatGPT, the latest #AI-based #chatbot system that has taken the internet by storm. https://
forbes.com/sites/bernardm
arr/2022/12/21/chatgpt-everything-you-really-need-to-know-in-simple-terms/
… -
Microsoft’s Structured Prompting Scales In-Context Learning to Thousands
By
–
Microsoft’s Structured Prompting Breaks In-Context Learning Length Limits, Scales to Thousands of Examples https://
syncedreview.com/2022/12/20/mic
rosofts-structured-prompting-breaks-in-context-learning-length-limits-scales-to-thousands-of-examples/
… -
Instruction Models Now Available in Open Source
By
–
Instruction models making their way into the open-source!
-
Fine-tuning vs. In-context Learning: Exploring Alternative Approaches
By
–
this is great question! i think its an unexplored space. finetuning is much more heavy weight (error prone, hard to do, more data required, etc), so it will be interesting to see if you can achieve similar performance just by using examples