Repository for our EMNLP 2023 paper "The Curious Case of Hallucinatory (Un)answerability: Finding Truths in the Hidden States of Over-Confident Large Language Models"

• Adjust prefix in unanswerability_env.yml to your Anaconda environment path.
• Run these commands:

conda env create -f unanswerability_env.yml
python -m spacy download en_core_web_sm
conda activate unanswerability_env

1. To download the dataset, run:

wget --load-cookies /tmp/cookies.txt "https://docs.google.com/uc?export=download&confirm=$(wget --quiet --save-cookies /tmp/cookies.txt --keep-session-cookies --no-check-certificate 'https://docs.google.com/uc?export=download&id=1q-6FIEGufKVBE3s6OdFoLWL2iHQPJh8h' -O- | sed -rn 's/.*confirm=([0-9A-Za-z_]+).*/\1\n/p')&id=1q-6FIEGufKVBE3s6OdFoLWL2iHQPJh8h" -O data.zip && rm -rf /tmp/cookies.txt

or directly download the file from Google Drive

2. uzip:

unzip raw_data.zip

To perform the zero-shot prompt-manipulation experiment, run:

python zero_shot_prompting.py --models <MODELS> --datasets <DATASETS> --return-only-generated-text --outdir /path/to/outdir

• <MODELS> - any one of 'Flan-UL2', 'Flan-T5-xxl', or 'OPT-IML' (can pass more than one).
• <DATASETS> - any one of 'squad', 'NQ', or 'musique' (can pass more than one).
• For prompt variants, add --prompt-variant <VARIANT_LIST>:
  • <VARIANT_LIST> - any one of 'variant1', 'variant2', 'variant3' (can pass more than one).
    • Default - 'variant1'.
• For development set experiments, add --devset.
• Output: Saves two .pt files in the specified outdir, one for answerable and one for un-answerable prompts.
  • Also saves the actual generated outputs in the subdir regular_decoding.

To perform the few-shot prompt-manipulation experiment, run:

python few_shot_prompting.py --models <MODELS> --datasets <DATASETS> --return-only-generated-text --outdir /path/to/outdir

• <MODELS> and <DATASETS> are similar to those in Zero-shot Prompting.
• Prompt variant can be changed like in Zero-shot Prompting.
• For in-context-learning examples variants - add --icl-examples-variant <ICL_VARIANT_LIST>:
  • <ICL_VARIANT_LIST> - any one of '1', '2', '3' (can pass more than one).

For beam relaxation experiments, just add --k-beams <BEAM_SIZE> to the Zero-shot Prompting command.

• Output: In addition to the subdir regular_decoding, an additional beam-relaxation subdir will be generated, with the beam-relaxed responses.

To evaluate the generated texts, run:

python -m evaluation.evaluate --indirs <INDIRS> --outdir /path/to/outdir 

• <INDIRS>: output directories from the prompting experiments.
• output: save under outdir:
  • QA-task-results.csv - results on the QA task for each prompt type (e.g., Regular-Prompt or Hint-Prompt).
  • unanswerability_classification_results.xlsx - unanswerability classification results for each prompt type.
• For results on development set, add --devset.

1. Generate Test Set Embeddings: Run the Zero-shot Prompting experiments without the --return-only-generated-text parameter.
   • This will also save the generations' embeddings (last hidden layer of first generated token) of the test set.
2. Generate Train Set Embeddings: In addition to step 1, also add --trainset.

• to run steps 1 and 2 on the first hidden layer of the first generated token, add --return-first-layer.
• Prompt variant can be changed like in Zero-shot Prompting.

Run:

python train_linear_classifiers.py --indir <INDIR> --outdir /path/to/outdir --dataset <DATASET> --prompt-type <PROMPT_TYPE> --epochs 100 --batch-size 16 --num-instances 1000

• <INDIR> - path to the directory with the saved embeddings (pt files) of the train set.
• <DATASET> - any one of 'squad', 'NQ', 'musique'.
• <PROMPT_TYPE> - 'Regular-Prompt' or 'Hint-Prompt'.
• To train a classifier on the first hidden layer of the first generated token, add --embedding-type first_hidden_embedding.
• output - save under "outdir//<EMBEDDING_TYPE>/<PROMPT_TYPE>/only_first_tkn/<MODEL_NAME>_1000N" the trained classifier.
  • <EMBEDDING_TYPE> - 'first_hidden_embedding' or 'last_hidden_embedding'.
  • <MODEL_NAME> - name of the model whose embeddings were used to train the classifier.

Run:

python evaluation/eval_linear_classifiers.py --indir <DATA_INDIR> --classifier-dir <CLASSIFIER_INDIR> --dataset <DATASET> --prompt-type <PROMPT_TYPE> --embedding-type <EMBEDDING_TYPE>

• <DATA_INDIR> - path to directory with the test set saved embeddings (pt files).
• <CLASSIFIER_INDIR> - path to the trained linear classifier.
• <DATASET> - any one of 'squad', 'NQ', 'musique' (should represent the dataset of the test set).
• <PROMPT_TYPE> - 'Regular-Prompt' or 'Hint-Prompt'.
• <EMBEDDING_TYPE> - 'first_hidden_embedding' or 'last_hidden_embedding'.

Run:

python figures_generation/PCA_plots_generation.py -i /path/to/folder/with/pt_files -o /path/to/outdir --prompt-type <PROMPT_TYPE> 

• <PROMPT_TYPE> - 'Regular-Prompt' or 'Hint-Prompt'.
• output - The generated 3-D PCA plots of the embedding space will be saved under "/path/to/outdir/last_hidden_embedding/only_first_tkn/<PROMPT_TYPE>".

1. Set Up Environment - Create a separate Conda environment for this experiment:
   • Adjust prefix in subspace_erasure.yml to your Anaconda environment path.
   • Run these commands:

conda env create -f subspace_erasure.yml
conda activate subspace_erasure

2. Make sure you have the embeddings of the train set from Preliminaries - Get Embeddings.

Run:

python train_concept_eraser.py --indir <INDIR> --outdir /path/to/outdir --dataset <DATASET> --prompt-type <PROMPT_TYPE> --epochs 500 --batch-size 16 --num-instances 1000

• <INDIR> - path to the directory with the embeddings (pt files) of the train set.
• <DATASET> - any one of 'squad', 'NQ', 'musique'
• <PROMPT_TYPE> - 'Regular-Prompt' or 'Hint-Prompt'
• output - trained eraser will be under "/path/to/outdir//<PROMPT_TYPE>"

Run:

python zero_shot_erasure_prompting.py --models <MODELS> --datasets <DATASETS> --outdir /path/to/outdir --eraser-dir /path/to/trained_eraser --only-first-decoding

• <MODELS> and <DATASETS> are similar to those in Zero-shot Prompting.
• Output: Saves two .pt files in the specified outdir, one for answerable and one for un-answerable prompts.
  • Also saves the actual generated outputs in the subdir regular_decoding.
• To evaluate the responses, follow the instructions under Evaluation.
• To visualize the embeddings, follow the instructions under Visualize Embedding Space.

If you use this in your work, please cite:

@inproceedings{slobodkin-etal-2023-curious,
    title = "The Curious Case of Hallucinatory (Un)answerability: Finding Truths in the Hidden States of Over-Confident Large Language Models",
    author = "Slobodkin, Aviv  and
      Goldman, Omer  and
      Caciularu, Avi  and
      Dagan, Ido  and
      Ravfogel, Shauli",
    editor = "Bouamor, Houda  and
      Pino, Juan  and
      Bali, Kalika",
    booktitle = "Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing",
    month = dec,
    year = "2023",
    address = "Singapore",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2023.emnlp-main.220",
    doi = "10.18653/v1/2023.emnlp-main.220",
    pages = "3607--3625",
    abstract = "Large language models (LLMs) have been shown to possess impressive capabilities, while also raising crucial concerns about the faithfulness of their responses. A primary issue arising in this context is the management of (un)answerable queries by LLMs, which often results in hallucinatory behavior due to overconfidence. In this paper, we explore the behavior of LLMs when presented with (un)answerable queries. We ask: do models \textit{represent} the fact that the question is (un)answerable when generating a hallucinatory answer? Our results show strong indications that such models encode the answerability of an input query, with the representation of the first decoded token often being a strong indicator. These findings shed new light on the spatial organization within the latent representations of LLMs, unveiling previously unexplored facets of these models. Moreover, they pave the way for the development of improved decoding techniques with better adherence to factual generation, particularly in scenarios where query (un)answerability is a concern.",
}