Sources of Hallucination by Large Language Models on Inference Tasks

Overview

This repository contains the code and data for the paper Sources of Hallucination by Large Language Models on Inference Tasks.

Data & Results Files

The data used in the paper can be found here, please download and put in the levyholt_files directory. The directory contains the following sub-folders:

• with_original: contains the directional Levy/Holt entries with their original context (i.e., the original entities as arguments).
• with_type: contains the directional Levy/Holt entries with their type context (i.e., with typed placeholders of the entities as arguments). These files should be processed with
• randprem_files: contains the random premise files used in Section 5/7 of the paper.
• swapped_entities: contains the files with entity-swapped entries used in the Section 6 of the paper.
• with_pseudoents_legacy: legacy files for a previous version of experiments with wugged words. These files are not used in the paper.

The model outputs for GPT models and LLaMA models are stored in the results directory. The directory contains the following sub-folders:

• gpt_results: contains the GPT model outputs (text-davinci-003 and gpt-4-0314) for the directional Levy/Holt entries; original, entity-swap, random-premise and hypothesis-only results are included for text-davinci-003.
• llama_results: contains the LLaMA model outputs (llama-65b) for the directional Levy/Holt entries, original, entity-swap, random-premise and hypothesis-only results are included.

Scripts

Running Models on Natural Language Inference Tasks

• gpt3_inference.py contains the code to acquire natural language inference results with GPT-3/4 models.
  • Example Script:
    • Dev set: python -u gpt3_inference.py --model_name text-davinci-003 --use_plhr [original/type/lowfreq/highfreq/randprem-orig/randprem-type] --split dev --in_context cot --num_templates 4
    • Test set: python -u gpt3_inference.py --model_name text-davinci-003 --use_plhr original --split dev --in_context cot --tplt_id 1 --num_templates 1
    • Hypothesis-only: python -u gpt3_inference.py --model_name text-davinci-003 --use_plhr original --split test --in_context lbl --num_templates 1 --hypothesis-only
• llama_inference.py contains the code for calculating the LLaMA results.
  • Example Script:
    • Dev set: python -u llama_inference.py --model_root [YOUR_PATH_TO_MODEL] --model_name llama-65b-hf --use_plhr [original/type/lowfreq/highfreq/randprem-orig/randprem-type] --split dev --in_context cot --task lh
    • Test set: python -u llama_inference.py --model_root [YOUR_PATH_TO_MODEL] --model_name llama-65b-hf --use_plhr original --split test --in_context cot --task lh --tplt_id 1
    • Hypothesis-only: python -u llama_inference.py --model_root [YOUR_PATH_TO_MODEL] --model_name llama-65b-hf --use_plhr original --split test --in_context lbl --task lh --single_statement h
  • Note: LLaMA is an LLM published by Meta, please visit their website to gain access to the model
• Attestation bias can be measured with the hyp-only setting in the two scripts above.
• Attestation results can be polled using the poll_attestation.py script.
• Relative frequency prior can be measured using the get_frequencies_ngram.py script.

Controlled Experiments

• randprem_experiments.py contains the code to calculate the probabilities of predicting Entail conditioned on attestation or relative frequency.
• attestation_controlled_experiment.py contains the code to calculate the model performance on attestation-consistent / adversarial subsets of Levy/Holt.
• frequency_controlled_experiments.py contains the code to calculate the model performance on frequency-consistent / adversarial subsets of Levy/Holt.