You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Large language models~(LLMs) are instruction followers, but it can bechallenging to find the best instruction for different situations, especiallyfor black-box LLMs on which backpropagation is forbidden. Instead of directlyoptimizing the discrete instruction, we optimize a low-dimensional soft promptapplied to an open-source LLM to generate the instruction for the black-boxLLM. On each iteration of the proposed method, which we call InstructZero, asoft prompt is converted into an instruction using the open-source LLM, whichis then submitted to the black-box LLM for zero-shot evaluation, and theperformance is sent to Bayesian optimization to produce new soft promptsimproving the zero-shot performance. We evaluate InstructZero on differentcombinations of open-source LLMs and APIs including Vicuna and ChatGPT. Ourresults show that InstructZero outperforms SOTA auto-instruction methods acrossa variety of downstream tasks. Our code and data are publicly available athttps://github.com/Lichang-Chen/InstructZero.
AkihikoWatanabe
changed the title
あ
InstructZero: Efficient Instruction Optimization for Black-Box Large
Language Models, Lichang Chen+, N/A, arXiv'23
Jun 16, 2023
URL
Affiliations
Abstract
Translation (by gpt-3.5-turbo)
Summary (by gpt-3.5-turbo)
The text was updated successfully, but these errors were encountered: