Buradasın

Promptbase: GPT-4 Prompt Engineering Resources

Yapay zekadan makale özeti

Core Components: Promptbase provides resources for best performing foundation models like GPT-4
Medprompt methodology combines three strategies for medical-level performance
Dynamic few-shot selection uses k-NN clustering for representative examples
Self-generated chain of thought enables complex reasoning
Choice-shuffle ensembling combines multiple algorithm outputs
Medprompt+ Extension: Achieved 89.1% score on MMLU benchmark without modifications
Scaling up to 20 ensemble calls improved performance to 89.56%
Added 10 direct few-shot prompts to classic Medprompt
GPT-4 helps determine appropriate prompting technique for each question
Performance improved by +0.5% across MMLU benchmark
Implementation Details: Available for multiple benchmarks including MMLU, HumanEval, and Math
Requires downloading and placing datasets in src/promptbase/datasets
Includes scripts for running various evaluation tests
Future plans include structured interviews and specialized tooling