Lookback Lens 🔎 🦙

Code for the paper "Detecting and Mitigating Contextual Hallucinations in Large Language Models Using Only Attention Maps"

Paper: https://arxiv.org/abs/2407.07071
Authors: Yung-Sung Chuang$^\dagger$, Linlu Qiu$^\dagger$, Cheng-Yu Hsieh$^\ddagger$, Ranjay Krishna$^\ddagger$, Yoon Kim$^\dagger$, James Glass$^\dagger$
$^\dagger$ Massachusetts Institute of Technology, $^\ddagger$ University of Washington

Introduction

When asked to summarize articles or answer questions given a passage, large language models (LLMs) hallucinate details and respond with unsubstantiated answers that are inaccurate with respect to the input context.

This paper describes a simple approach for detecting such contextual hallucinations. We hypothesize that contextual hallucinations are related to the extent to which an LLM attends to information in the provided context versus its own generations. Based on this intuition, we propose a simple hallucination detection model whose input features are given by the ratio of attention weights on the context versus newly generated tokens (for each attention head). We find that a linear classifier based on these lookback ratio features is as effective as a richer detector that utilizes the entire hidden states of an LLM or a text-based entailment model.

The lookback ratio-based detector—Lookback Lens—is found to transfer across tasks and even models, allowing a detector that is trained on a 7B model to be applied (without retraining) to a larger 13B model.

We further apply this detector to mitigate hallucination generations, and find that a simple classifier-guided sampling approach is able to reduce the amount of hallucinations. For example, the detector is able to reduce hallucinations by 9.6% in the XSum summarization task.

Installation

Python version: 3.9.5
CUDA toolkit version: 11.7

pip install -r requirements.txt
pip install -e ./transformers-4.32.0

gzip -d data/nq-open-10_total_documents_gold_at_4.jsonl.gz

Preparation 📚

**Hint: Skip step 01 & 02 by downloading the precomputed lookback ratios & annotations here.**

Step 01: Extracting Lookback Ratios from Attention Weights (NQ and CNN/DM) (Optional)

To load LLaMA2 models/tokenizers, please login with huggingface-cli login, or add the argument --auth_token <hf_auth_token> where <hf_auth_token> is your huggingface auth token with LLaMA2 access.

python step01_extract_attns.py --model-name meta-llama/Llama-2-7b-chat-hf --data-path data/nq-open-10_total_documents_gold_at_4.jsonl --output-path lookback-ratio-nq-7b.pt
python step01_extract_attns.py --model-name meta-llama/Llama-2-7b-chat-hf --data-path data/cnndm-1000.jsonl --output-path lookback-ratio-cnndm-7b.pt

Step 02: Run GPT-4o Annotation (NQ and CNN/DM) (Optional)

OPENAI_API_KEY={your_key} python step02_eval_gpt4o.py --hyp lookback-ratio-nq-7b.pt --ref data/nq-open-10_total_documents_gold_at_4.jsonl --out anno-nq-7b.jsonl
OPENAI_API_KEY={your_key} python step02_eval_gpt4o.py --hyp lookback-ratio-cnndm-7b.pt --ref data/xsum-1000.jsonl --out anno-cnndm-7b.jsonl

Logistic Regression Classifiers (Lookback Lens) 📈

Step 03: Fitting Lookback Lens Classifiers (NQ and CNN/DM)

To load LLaMA2 models/tokenizers, please login with huggingface-cli login, or add the argument --auth_token <hf_auth_token> where <hf_auth_token> is your huggingface auth token with LLaMA2 access.

# Predefined Span
python step03_lookback_lens.py --anno_1 anno-nq-7b.jsonl --anno_2 anno-cnndm-7b.jsonl --lookback_ratio_1 lookback-ratio-nq-7b.pt --lookback_ratio_2 lookback-ratio-cnndm-7b.pt
# Sliding Window (=8)
python step03_lookback_lens.py --anno_1 anno-nq-7b.jsonl --anno_2 anno-cnndm-7b.jsonl --lookback_ratio_1 lookback-ratio-nq-7b.pt --lookback_ratio_2 lookback-ratio-cnndm-7b.pt --sliding_window 8

The output will be similar to:

# Predefined Span

======== Results:
                  , Train AUROC (on A), Test AUROC (on A), Transfer AUROC (on B)
A=nq-7b;B=cnndm-7b, 0.9867235784623354, 0.9140908050233869, 0.8526936562673579
A=cnndm-7b;B=nq-7b, 0.9844307377081996, 0.8720309189629751, 0.8203155443540785

# Sliding Window (=8)

======== Results:
                  , Train AUROC (on A), Test AUROC (on A), Transfer AUROC (on B)
A=nq-7b;B=cnndm-7b, 0.8858071459740011, 0.8663781955546325, 0.6624004639123215
A=cnndm-7b;B=nq-7b, 0.8650978795284527, 0.8474340844981891, 0.6608756591251488

Inference 🏃

Step 04: Run Greedy vs Classifier Guided Decoding (NQ and XSum)

We perform decoding with classifiers/classifier_anno-cnndm-7b_sliding_window_8.pkl for both tasks to test the in-domain (XSum) and out-of-domain (NQ) performance of the Lookback Lens Guided Decoding.

To load LLaMA2 models/tokenizers, please login with huggingface-cli login, or add the argument --auth_token <hf_auth_token> where <hf_auth_token> is your huggingface auth token with LLaMA2 access.

# Greedy (NQ)
python step04_run_decoding.py --model_name meta-llama/Llama-2-7b-chat-hf/ --data_path data/nq-open-10_total_documents_gold_at_4.jsonl --output_path output-nq-open-greedy-decoding.jsonl --num_gpus 1
# Lookback Lens Guided Decoding (NQ)
python step04_run_decoding.py --model_name meta-llama/Llama-2-7b-chat-hf/ --data_path data/nq-open-10_total_documents_gold_at_4.jsonl --output_path output-nq-open-lookback-decoding.jsonl --num_gpus 1 --do_sample --guiding_classifier classifiers/classifier_anno-cnndm-7b_sliding_window_8.pkl --chunk_size 8 --num_candidates 8 

# Greedy (XSum)
python step04_run_decoding.py --model_name meta-llama/Llama-2-7b-chat-hf/ --data_path data/xsum-1000.jsonl --output_path output-xsum-greedy-decoding.jsonl --num_gpus 1
# Lookback Lens Guided Decoding (XSum)
python step04_run_decoding.py --model_name meta-llama/Llama-2-7b-chat-hf/ --data_path data/xsum-1000.jsonl --output_path output-xsum-lookback-decoding.jsonl --num_gpus 1 --do_sample --guiding_classifier classifiers/classifier_anno-cnndm-7b_sliding_window_8.pkl --chunk_size 8 --num_candidates 8 

If too slow: Parallel (Sharded) Inference

Running inference in sharded mode can be done by setting --parallel --total_shard 4 --shard_id 0 for the first shard, --parallel --total_shard 4 --shard_id 1 for the second shard, and so on. The dataset will be split into 4 shards and the inference of each shard can be run in parallel.

Evaluation 📊

Run Exact Match Evaluation (NQ)

python eval_exact_match.py --hyp output-nq-open-greedy-decoding.jsonl --ref data/nq-open-10_total_documents_gold_at_4.jsonl
python eval_exact_match.py --hyp output-nq-open-lookback-decoding.jsonl --ref data/nq-open-10_total_documents_gold_at_4.jsonl

The output will be similar to:

# Greedy
Best span EM: 0.711864406779661
# Lookback Lens Guided Decoding
Best span EM: 0.7419962335216572 (by random sampling so the result may vary)

Run GPT-4o Evaluation (XSum)

OPENAI_API_KEY={your_key} python step02_eval_gpt4o.py --hyp output-xsum-greedy-decoding.jsonl --ref data/xsum-1000.jsonl --out record-gpt4o-eval-xsum-greedy-decoding.jsonl 
OPENAI_API_KEY={your_key} python step02_eval_gpt4o.py --hyp output-xsum-lookback-decoding.jsonl --ref data/xsum-1000.jsonl --out record-gpt4o-eval-xsum-lookback-decoding.jsonl 

The output will be similar to:

# Greedy
Accuracy: 0.490
# Lookback Lens Guided Decoding
Accuracy: 0.586
(the result may vary due to the randomness of GPT-4o API and the randomness of sampling)

Citation

Please cite our paper if it's helpful to your work!

@article{chuang2024lookback,
  title={Lookback Lens: Detecting and Mitigating Contextual Hallucinations in Large Language Models Using Only Attention Maps},
  author={Chuang, Yung-Sung and Qiu, Linlu and Hsieh, Cheng-Yu and Krishna, Ranjay and Kim, Yoon and Glass, James},
  journal={arXiv preprint arXiv:2407.07071},
  year={2024},
}