Skip to content

Latest commit

 

History

History
15 lines (10 loc) · 889 Bytes

File metadata and controls

15 lines (10 loc) · 889 Bytes

LLM experimental code for Adaptive Learn-then-Test: Statistically Valid and Efficient Hyperparameter Selection

This code is built based on https://github.com/keirp/automatic_prompt_engineer

  • aLTT is implemented at /automatic_prompt_engineer/altogether_ape.py

  • Code to run instruction induction experiment can be found at /experiments/main.py

  • Example execution of the code for the target risk alpha and tolerance delta reads:

python main.py --risk_control_mode ['FWER' or 'FDR'] --alpha $alpha$ --delta $delta$
  • Generated prompts using Llama3.3 70B (entire candidate set) as well as corresponding loss table obtained from Llama3 8B Instruct can be found in experiments/cache_original while 20 different random split of validation and testing data can be found in experiments/cache
  • Saved results with plot-generating code can be found in /plotting_with_saved_results/