Strength in Numbers: Estimating Confidence of Large Language Models by Prompt Agreement

Gwenyth Portillo Wightman and Alexandra DeLucia and Mark Dredze

This repository contains information for replicating the approach used in Strength in Numbers: Estimating Confidence of Large Language Models by Prompt Agreement.

Confidence Estimates

We provide code to compute swapped pairs and expected calibration error (ECE) in ./src/. Both scripts provide the option to compute swapped pairs and ECE using either the rand index (rand) or majority vote (majority) strategies, discussed in the paper.

ECE and swapped pairs results are saved to ./data/statistics/.

Example command for running compute_swapped_pairs.py for the T0++ model using the majority vote strategy:

python compute_swapped_pairs.py \  
  --OIP_input_dir '../../data/predictions/OIP_prompt/T0/cos_e' \  
  --T0_prompt_input_dir '../../data/predictions/T0_prompts/T0/cos_e' \  
  --paraphrase_input_dir '../../data/predictions/paraphrase/T0/cos_e' \  
  --output_dir '../../data/statistics/swapped_pairs/majority' \  
  --model_name T0 \  
  --probability_col log_probabilities \  
  --sorting_algorithm mergesort \  
  --rank_by_rand_or_majority majority

Example command for running compute_ece.py for the T0++ model using the rand index strategy:

python compute_ece.py \  
  --OIP_input_dir '../../data/predictions/OIP_prompt/T0/cos_e' \  
  --T0_prompt_input_dir '../../data/predictions/T0_prompts/T0/cos_e' \  
  --paraphrase_input_dir '../../data/predictions/paraphrase/T0/cos_e' \  
  --output_dir '../../data/statistics/ece' \  
  --model_name T0 \  
  --probability_col log_probabilities \  
  --sorting_algorithm mergesort \  
  --rank_by_rand_or_majority rand 

Prediction Data

For demonstration purposes, we provide examples of the JSONL data files that compute_ece.py and compute_swapped_pairs.py expect in ./data/predictions/. We include data for the three kinds of prompts:

• OIP_prompt: the dataset's originally intended prompt, which we refer to as the Single, Human Prompt prompt in the paper,
• T0_prompts: the prompts written for the dataset when Sanh et al. (2021) trained T0++, which we refer to as the Multiple, Human Prompts in the paper,
• paraphrase: the prompts that we automatically generate according to the method in the paper; we refer to these as the Automatically Generated Prompts in the paper.

Note that the prediction data provided in this repository is a sample of the predictions used in the paper. We provide a small set of predictions from T0++, GPT davinci, and Flan-T5-XXL for one dataset (cos_e/v1.11), for a subset of inputs from the dataset, and a subset of prompts.

Output

ECE and swapped pairs results are saved to ./data/statistics/ece/.

An example output CSV for ECE using T0++ and the rand index strategy is found at ./data/statistics/ece/rand/summary/T0.csv:

Dataset,OIP ECE,T0 Prompt ECE,Paraphrase ECE,T0 Prompt ECE < OIP ECE,Paraphrase ECE < OIP ECE,Paraphrase ECE < T0 Prompt ECE
cos_e,2.4400360420346257,1.05856421739057,12.591883913675945,True,False,False

If compute_ece.py for T0++ and rand had been run on multiple datasets, then multiple datasets would be listed in the CSV.

References

If you use the materials in this repository, please use the following citation:

@inproceedings{portillowightman2023strength,
    title={Strength in Numbers: Estimating Confidence of Large Language Models by Prompt Agreement},
    author={Portillo Wightman, Gwenyth and DeLucia, Alexandra and Dredze, Mark},
    booktitle={Proceedings of the 3rd Workshop on Trustworthy Natural Language Processing (TrustNLP 2023)},
    pages={326--362},
    year={2023}
}