nlp.lmd 0.2.0

Language Model Decomposition

lmd

Code for paper titled "Language Model Decomposition: Quantifying the Dependency and Correlation of Language Models" (accepted EMNLP 2022). The arxiv version is here:

Install

Create virtual env if needed

python3 -m venv .venv
source .venv/bin/activate

Install from pip

pip install nlp.lmd

Install from source

git clone [e-mail]:haozhg/lmd.git
cd lmd
pip install -e .

To use lmd cli, run lmd --help or python -m lmd.cli --help

$ lmd --help
usage: Language Model Decomposition [-h] [--target TARGET] [--basis BASIS]
                                    [--tokenizer-name TOKENIZER_NAME]
                                    [--max-seq-length MAX_SEQ_LENGTH]
                                    [--batch-size BATCH_SIZE]
                                    [--dataset-name DATASET_NAME]
                                    [--dataset-config-name DATASET_CONFIG_NAME]
                                    [--val-split-percentage VAL_SPLIT_PERCENTAGE]
                                    [--test-split-percentage TEST_SPLIT_PERCENTAGE]
                                    [--max-train-samples MAX_TRAIN_SAMPLES]
                                    [--max-val-samples MAX_VAL_SAMPLES]
                                    [--max-test-samples MAX_TEST_SAMPLES]
                                    [--preprocessing-num-workers PREPROCESSING_NUM_WORKERS]
                                    [--overwrite_cache OVERWRITE_CACHE]
                                    [--preprocess-dir PREPROCESS_DIR]
                                    [--embedding-dir EMBEDDING_DIR]
                                    [--results-dir RESULTS_DIR]
                                    [--models-dir MODELS_DIR] [--alpha ALPHA]
                                    [--log-level LOG_LEVEL]
                                    [--try-models TRY_MODELS]
                                    [--pre-select-multiplier PRE_SELECT_MULTIPLIER]
                                    [--seed SEED]

Results

To reproduce the results in Appendix B of the paper, run bash scripts/run.sh. The results are also stored in results/128k