This is codebase for experiments run doing text-based active learning using large language models as generative models for query synthesis. Contact [email protected] for more details.
- Clone the repository:
git clone [email protected]:OxAI-Safety-Hub/al-llm-experiments.git
- Optional (but recommended): create a new virtual environment.
- Install the requirements:
pip install -r requirements.txt
- Set up Weights and Biases
- Change the Weights and Biases entity name in
/al_llm/constants.py.
- The main script for running experiments is
scripts/run_experiment.py. - First install the
al_llmpackage locally in editable mode:
pip install -e .
- Pass the
--helpflag to see the list of options:
python /scripts/run_experiment.py --help
- The first argument is the run ID. A good convention for run ID is:
{DATASET}_{CLASSIFIER}_{SAMPLE_GENERATOR}_{ACQUISTION_FUNCTION}_{NUMBER}
using abbreviations. For example, the second pool-based experiment with the Rotten Tomatoes dataset, which uses a plain classifier and max uncertainty acquisition function might be called:
rt_plain_pool_mu_2
If there are special features you can add them at the end, before the number.
- By default the experiment runs in the 'Experiments' project. To change this specify the
--project-nameoption. - To select the GPU, use something like
--cuda-device 'cuda:1'. The default is to use the 0th device. - By default, we run a single experiment for one seed. To run multiple of the same experiment over different seeds, add the
--multiple-seedsflag. - In terms of configuring the experiment parameters, you'll most likely want to play around with the following options:
--dataset-name
--classifier-base-model
--use-tapted-classifier
--sample-generator-base-model
--use-tapted-sample-generator
--sample-generator-temperature
--sample-generator-top-k
--acquisition-function
But other options may be interesting.
The following guides are located in the docs folder.