This is the repository for the paper "Generalists vs. Specialists: Evaluating LLMs on Highly-Constrained Biophysical Sequence Optimization Tasks," by Angelica Chen^*, Samuel D. Stanton^*, Frances Ding, Robert G. Alberstein, Andrew M. Watkins, Richard Bonneau, Vladimir Gligorijević, Kyunghyun Cho, Nathan C. Frey

1. Clone this repo to your home directory! (Some of the scripts in this repo assume that your code is stored in ~/llome. If the path to your repo is different, change the path_to_repo option in the hydra config.)
2. Install the requirements in a virtual environment:

uv venv --python 3.10
source .venv/bin/activate
uv pip install -r requirements.txt

1. Run aws configure to ensure you are configured for access to S3 (if using S3 storage).

python -m run_pipeline --config-name=pipeline_sanity_check_marge local_output_dir=<PATH_TO_LOCAL_OUTPUT_DIR> parent_output_dir=<S3_OUTPUT_DIR>

• Storage: This pipeline makes use of both local storage (local_output_dir in the YAML config) and S3 storage (parent_output_dir in the YAML config). All model checkpoints are automatically copied to S3 and deleted from local storage after the end of a training job. By default, all generated seeds, generations, and training data are stored in <config.parent_output_dir>/<config.run_name>, where config is the Hydra config.
  • To disable S3 storage and use only local storage, simply set parent_output_dir: "null" in the config.
• Slurm Configurations: All training, dataset generation/formatting, and generation jobs are launched as separate SLURM jobs. These SLURM jobs are easily configured with the respective slurm_args subsections in the config.
• Resuming jobs: This pipeline is designed to automatically resume previous runs, assuming that the pipeline is run with the same arguments as a prior run. To disable this behavior, use --overwrite=True.
• Training large models: We currently use DDP (without accompanying model parallelism) to train our models (FSDP is a TODO though!), which means that you likely need GPU RAM that is ~4x the size of the model to train in full precision. Our code is currently only tested on single-node multi-GPU setups.
• Analyzing results: To assess the performance of the LLM at iteration $i$, analyze the training data generated for iteration $i+1$ -- this accounts for both the likelihood-based ranking and filtering of the LLM at iteration $i$.

Note: Update the slurm_args section in each config before running your experiment!

LLOME-SFT:

python -m run_pipeline --config-name=pipeline_sft_f2 local_output_dir=<PATH_TO_LOCAL_OUTPUT_DIR> parent_output_dir=<S3_OUTPUT_DIR>

LLOME-MargE:

python -m run_pipeline --config-name=pipeline_marge_f2 local_output_dir=<PATH_TO_LOCAL_OUTPUT_DIR> parent_output_dir=<S3_OUTPUT_DIR>

LLOME-DPO:

python -m run_pipeline --config-name=pipeline_dpo_f2 local_output_dir=<PATH_TO_LOCAL_OUTPUT_DIR> parent_output_dir=<S3_OUTPUT_DIR>