Toward Understanding Hallucinations through Latent Knowledge

Paper Link

This is a Repository of all the Experiements conducted to gather more deep understanding into Extraction of Latent Knowledge by Using a particular technique of Contrast Consistent Search

The Study can be Segmented into the following Categories

• Reimplementation of the Paper that inspired this project. Burns et al
• Pilot Study on Vision Language models and Analysis of POPE Dataset Code
• Extension of CCS in Vision-Language Models Folder
• Contrastive Study design between Performance of Base model Vs CCS on CLIP and LLaVa 1.5B Code
• Validating the Robustness of CCS by Conducting a study on Text Transfer and Adversarial Robustness Code

Prerequisites

First install the Dependencies

Note: It is advised to create a virtual environment using either conda or venv

pip install -r requirements.txt

---

Reimplementation of the paper

In the LLM reimplementation part, we tested CCS on every layer of LLama-3.2 (instruct and base version) Qwen-2.5 (instruct and base version) using different databases. Compared to the original code, main modifications are made to LLM_reimplementation/utils.get_individual_hidden_states, LLM_reimplementation/utils.get_dataloader, LLM_reimplementation/utils.ContrastDataset

Args used

Model: ['Qwen/Qwen2.5-3B-Instruct','meta-llama/Llama-3.2-3B-Instruct'] 

Dataset: ['imdb','amazon_polarity','ag_news','dbpedia_14','multi_nli',["glue","qnli"],"gimmaru/story_cloze-2016","piqa"]

Split: ['test','test','test','test','test','validation_mismatched','test','train']

Prompt: ['imdb','amazon_polarity','ag_news','dbpedia_14','multi_nli',["glue","qnli"],"story_cloze/2016","piqa"]

Running the Code

1. To generate the hidden layers embedding on one model and one dataset, please run:

python LLM_reimplementation/generate.py --model_name [] --dataset_name [] --split [] --prompt_name [] --all_layers

2. Given a list of models and a list of datasets and splits, using the hidden embedding already gotten to get CCS performance for each model, each layer, each dataset.

python LLM_reimplementation/evaluation.py --models [] --datasets [] --splits [] --save_dir [] --all_layers

Use any of the Arguments for dataset_name, prompt_name_

3. Code for ploting can be run using :

python LLM_reimplementation/plot.py

Pilot Study on Vison Language models and Analysis of POPE Dataset

Running the Code

All the .ipynb files can be run as is, it includes all the dependencies that need to be installed and are plug and play.

Extension of CCS in Vision-Language Models

In the VLM_hallucination part, we studied the hallucination problem using LLaVa-1.5-7b-hf model, and POPE dataset.

Running the Code

1. We use VLM_hallucination/VLM_hallucination_answer_generate.py to generate model answers to POPE dataset, as well as the logits value of answers. Please run:

python VLM_hallucination/VLM_hallucination_answer_generate.py --dataset_name lmms-lab/POPE default --save_logits_dir []

2. After we get the hallucination dataset, we generate the hidden embeddings of hallucination contrast pairs. Please run:

python VLM_hallucination/VLM_hallucination_pair_generate.py --dataset_name lmms-lab/POPE default --save_logits_dir [] --save_hidden_dir []

3. After we get the hidden embedding, we test CCS's performance as a hallucination classifier. Please run:

python VLM_hallucination/evaluate.py --dataset_name lmms-lab/POPE default --save_hidden_dir []

4. To construct truth contrast pair for POPE and save the hidden embeddings, Please run:

python VLM_hallucination/VLM_truth_pair_generate.py --dataset_name lmms-lab/POPE default --save_hidden_dir []

4.1 To save hallucination CCS, please run:

python VLM_hallucination/evaluate_save_probe.py --save_hidden_dir [] --save_probe_dir []

4.2 To test the truth contrast pair on the hallucination CCS, please run:

python VLM_hallucination/evaluate_load_probe.py --save_hidden_dir [] --save_probe_dir []

Contrastive Study design between Performance of Base model Vs CCS on CLIP and LLaVa 1.5B Using the Contrast pair Questions

Running the Code

All the .ipynb files can be run as is, it includes all the dependencies that need to be installed and are plug and play.

###Validating the Robustness of CCS by Conducting a study on Text Transfer and Adversarial Robustness

Running the Code

All the .ipynb files can be run as is, it includes all the dependencies that need to be installed and are plug and play.