thuanystuart / DD3412-chain-of-verification-reproduction

Re-implementation of the paper "Chain-of-Verification Reduces Hallucination in Large Language Models" for hallucination reduction. Developed as a final project of the Advanced Deep Learning course (DD3412) at KTH.

DD3412-chain-of-verification-reproduction

🔥 Quickstart

First, create a python virtual environment and install the requirements:

conda create -n CoVe python=3.10
conda activate CoVe

To install requirements:

pip install -r requirements.txt

Then, create a file named .configuration to store the confidential data, such as OpenAI and HuggingFace API keys. This file should look like:

OPENAI_API_KEY=sk-abc.....
HF_API_KEY=hf_abc.......

💻 Run the experiments

🎯 Running with Bash Script

To execute CoVe, use the command below:

>> bash scripts/wikidata.sh # for wikidata task, and Llama2-70b model

• Available models are: llama2, llama2_70b, llama-65b, gpt3
• Available settings are: joint, two_step, factored
• Available tasks are: wikidata, wikidata_category, multispanqa

📦 Running with Apptainer

You can run the experiments using Apptainer with the following steps:

1. Build the Apptainer container by utilizing the provided my_apptainer.def file:

# Build the Apptainer container
apptainer build my_container.sif my_apptainer.def

2. Run the CoVe experiments within the Apptainer container by executing the apptainer_job:

SBATCH apptainer_job.sh

Alternatively, you can use the following command:

apptainer exec --nv my_apptainer.sif ./scripts/wikidata.sh # for Wikidata task and Llama2-70b