JMTEB is a benchmark for evaluating Japanese text embedding models. It consists of 5 tasks.
This is an easy-to-use evaluation script designed for JMTEB evaluation.
JMTEB leaderboard is here. If you would like to submit your model, please refer to the submission guideline.
git clone [email protected]:sbintuitions/JMTEB
cd JMTEB
poetry install
poetry run pytest testsThe following command evaluate the specified model on the all the tasks in JMTEB.
poetry run python -m jmteb \
--embedder SentenceBertEmbedder \
--embedder.model_name_or_path "<model_name_or_path>" \
--save_dir "output/<model_name_or_path>"Note
In order to gurantee the robustness of evaluation, a validation dataset is mandatorily required for hyperparameter tuning. For a dataset that doesn't have a validation set, we set the validation set the same as the test set.
By default, the evaluation tasks are read from src/jmteb/configs/jmteb.jsonnet.
If you want to evaluate the model on a specific task, you can specify the task via --evaluators option with the task config.
poetry run python -m jmteb \
--evaluators "src/configs/tasks/jsts.jsonnet" \
--embedder SentenceBertEmbedder \
--embedder.model_name_or_path "<model_name_or_path>" \
--save_dir "output/<model_name_or_path>"Note
Some tasks (e.g., AmazonReviewClassification in classification, JAQKET and Mr.TyDi-ja in retrieval, esci in reranking) are time-consuming and memory-consuming. Heavy retrieval tasks take hours to encode the large corpus, and use much memory for the storage of such vectors. If you want to exclude them, add --eval_exclude "['amazon_review_classification', 'mrtydi', 'jaqket', 'esci']". Similarly, you can also use --eval_include to include only evaluation datasets you want.
Note
If you want to log model predictions to further analyze the performance of your model, you may want to use --log_predictions true to enable all evaluators to log predictions. It is also available to set whether to log in the config of evaluators.
There are two ways to enable multi-GPU evaluation.
- New class
DataParallelSentenceBertEmbedder(here).
poetry run python -m jmteb \
--evaluators "src/configs/tasks/jsts.jsonnet" \
--embedder DataParallelSentenceBertEmbedder \
--embedder.model_name_or_path "<model_name_or_path>" \
--save_dir "output/<model_name_or_path>"- With
torchrun, multi-GPU inTransformersEmbedderis available. For example,
MODEL_NAME=<model_name_or_path>
MODEL_KWARGS="\{\'torch_dtype\':\'torch.bfloat16\'\}"
torchrun \
--nproc_per_node=$GPUS_PER_NODE --nnodes=1 \
src/jmteb/__main__.py --embedder TransformersEmbedder \
--embedder.model_name_or_path ${MODEL_NAME} \
--embedder.pooling_mode cls \
--embedder.batch_size 4096 \
--embedder.model_kwargs ${MODEL_KWARGS} \
--embedder.max_seq_length 512 \
--save_dir "output/${MODEL_NAME}" \
--evaluators src/jmteb/configs/jmteb.jsonnetNote that the batch size here is global batch size (per_device_batch_size × n_gpu).