Experiments for efficient repurposing embedding models via contrastive learning of pre-trained LLMs.

• Python 3.10 -- 3.11

python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt
pip install -e .

The shell scripts for downloading and light pre-processing of various datasets are located in scripts/data_prep. E.g., to download NLI dataset run:

./scripts/data_prep/nli.sh

and the dataset in JSONL file will be save to data/nli/all.jsonl.

The data used for experiments in the paper (the BAAI data set) can be downloaded from https://data.baai.ac.cn/details/BAAI-MTP (login required).

To run training and MTEB evaluation for the four methods considered in the paper (full fine-tuning, layer freezing, only-bias tuning, and LoRA), please run an appropriate script from among experiments/*.py. For instance:

./experiments/train_and_eval_freezing.py

The checkpoints of the model and other data will be saved to the result directory.

experiments/*.gin files associated with the Python scripts define configurations of the experiments (learning rate, weight decay, LoRA rank, etc.) You can modify them.

In the file data/mteb/test_tasks_list there is a short list of MTEB tasks used for evaluation for debugging purposes. The list used for evaluation in the paper is in data/mteb/broad_tasks_list. (The former list is specified in the config files.)