Implicit meta-learning may lead language models to trust more reliable sources

This repository contains code for the language model experiments from the paper Implicit meta-learning may lead language models to trust more reliable sources (paper, ICML 2024 poster).

Steps to get started:

1. Clone the repository

git clone https://github.com/krasheninnikov/internalization.git
cd internalization

2. Configure your Python environment

• Step 1. Create and activate a new Conda environment:

  conda create --name internalization python=3.11
  conda activate internalization

• Step 2. Install the dependencies and download the datasets:

  pip install -r requirements.txt
  # download the datasets from Google Drive
  gdown --folder 'https://drive.google.com/drive/folders/1KQDClI3cbFzPhzfknF2xmtqE-aIW1EDf?usp=sharing'

• Step 3 (Optional). Configure wandb:

  wandb login
  wandb init --entity=your-entity --project=your-project

3. Run the experiment

To run the experiment with the default configuration (configs/current_experiment.yaml), use the following command:

python -m src.run

Choosing/modifying/creating an experiment configuration. Go to the configs directory to select an existing configuration or create a new one. Some parameter descriptions can be found in the configs readme.

Once the configuration is ready, run the experiment with the following command:

python -m src.run -cp <your-config-path>