Awesome

Installation

pip install mteb

Example Usage

• Using a Python script:

import mteb
from sentence_transformers import SentenceTransformer

# Define the sentence-transformers model name
model_name = "average_word_embeddings_komninos"

model = mteb.get_model(model_name) # if the model is not implemented in MTEB it will be eq. to SentenceTransformer(model_name)
tasks = mteb.get_tasks(tasks=["Banking77Classification"])
evaluation = mteb.MTEB(tasks=tasks)
results = evaluation.run(model, output_folder=f"results/{model_name}")

Prompts can be passed to the SentenceTransformer model using the prompts parameter. The following code shows how to use prompts with SentenceTransformer:

from sentence_transformers import SentenceTransformer


model = SentenceTransformer("average_word_embeddings_komninos", prompts={"query": "Query:", "passage": "Passage:"})
evaluation = mteb.MTEB(tasks=tasks)

In prompts the key can be:

1. Prompt types (passage, query) - they will be used in reranking and retrieval tasks
2. Task type - these prompts will be used in all tasks of the given type
   1. BitextMining
   2. Classification
   3. MultilabelClassification
   4. Clustering
   5. PairClassification
   6. Reranking
   7. Retrieval
   8. STS
   9. Summarization
   10. InstructionRetrieval
3. Pair of task type and prompt type like Retrival-query - these prompts will be used in all classification tasks
4. Task name - these prompts will be used in the specific task
5. Pair of task name and prompt type like NFCorpus-query - these prompts will be used in the specific task

• Using CLI

mteb available_tasks

mteb run -m sentence-transformers/all-MiniLM-L6-v2 \
    -t Banking77Classification  \
    --verbosity 3

# if nothing is specified default to saving the results in the results/{model_name} folder

• Using multiple GPUs in parallel can be done by just having a custom encode function that distributes the inputs to multiple GPUs like e.g. here or here.

Usage Documentation

Click on each section below to see the details.

Task selection

Tasks can be selected by providing the list of datasets, but also

• by their task (e.g. "Clustering" or "Classification")

tasks = mteb.get_tasks(task_types=["Clustering", "Retrieval"]) # Only select clustering and retrieval tasks

• by their categories e.g. "s2s" (sentence to sentence) or "p2p" (paragraph to paragraph)

tasks = mteb.get_tasks(categories=["s2s", "p2p"]) # Only select sentence2sentence and paragraph2paragraph datasets

• by their languages

tasks = mteb.get_tasks(languages=["eng", "deu"]) # Only select datasets which contain "eng" or "deu" (iso 639-3 codes)

You can also specify which languages to load for multilingual/cross-lingual tasks like below:

import mteb

tasks = [
    mteb.get_task("AmazonReviewsClassification", languages = ["eng", "fra"]),
    mteb.get_task("BUCCBitextMining", languages = ["deu"]), # all subsets containing "deu"
]

# or you can select specific huggingface subsets like this:
from mteb.tasks import AmazonReviewsClassification, BUCCBitextMining

evaluation = mteb.MTEB(tasks=[
        AmazonReviewsClassification(hf_subsets=["en", "fr"]) # Only load "en" and "fr" subsets of Amazon Reviews
        BUCCBitextMining(hf_subsets=["de-en"]), # Only load "de-en" subset of BUCC
])
# for an example of a HF subset see "Subset" in the dataset viewer at: https://huggingface.co/datasets/mteb/bucc-bitext-mining

Running a Benchmark

mteb comes with a set of predefined benchmarks. These can be fetched using get_benchmark and run in a similar fashion to other sets of tasks. For instance to select the 56 English datasets that form the "Overall MTEB English leaderboard":

import mteb
benchmark = mteb.get_benchmark("MTEB(eng, classic)")
evaluation = mteb.MTEB(tasks=benchmark)

The benchmark specified not only a list of tasks, but also what splits and language to run on. To get an overview of all available benchmarks simply run:

import mteb
benchmarks = mteb.get_benchmarks()

Generally we use the naming scheme for benchmarks MTEB(*), where the "*" denotes the target of the benchmark. In the case of a language, we use the three-letter language code. For large groups of languages, we use the group notation, e.g., MTEB(Scandinavian) for Scandinavian languages. External benchmarks implemented in MTEB like CoIR use their original name. When using a benchmark from MTEB please cite mteb along with the citations of the benchmark which you can access using:

benchmark.citation

Passing in encode arguments

To pass in arguments to the model's encode function, you can use the encode keyword arguments (encode_kwargs):

evaluation.run(model, encode_kwargs={"batch_size": 32})

Selecting evaluation split

You can evaluate only on test splits of all tasks by doing the following:

evaluation.run(model, eval_splits=["test"])

Note that the public leaderboard uses the test splits for all datasets except MSMARCO, where the "dev" split is used.

Selecting evaluation subset

You can evaluate only on selected subsets. For example, if you want to evaluate only the subset_name_to_run subset of all tasks, do the following:

evaluation.run(model, eval_subsets=["subset_name_to_run"])

Monolingual tasks have default subset, other tasks have subsets that are specific to the dataset.

Using a custom model

Models should implement the following interface, implementing an encode function taking as inputs a list of sentences, and returning a list of embeddings (embeddings can be np.array, torch.tensor, etc.). For inspiration, you can look at the mteb/mtebscripts repo used for running diverse models via SLURM scripts for the paper.

import mteb
from mteb.encoder_interface import PromptType
import numpy as np


class CustomModel:
    def encode(
        self,
        sentences: list[str],
        task_name: str,
        prompt_type: PromptType | None = None,
        **kwargs,
    ) -> np.ndarray:
        """Encodes the given sentences using the encoder.
        
        Args:
            sentences: The sentences to encode.
            task_name: The name of the task.
            prompt_type: The prompt type to use.
            **kwargs: Additional arguments to pass to the encoder.
            
        Returns:
            The encoded sentences.
        """
        pass

model = CustomModel()
tasks = mteb.get_tasks(tasks=["Banking77Classification"])
evaluation = MTEB(tasks=tasks)
evaluation.run(model)

Evaluating on a custom dataset

To evaluate on a custom task, you can run the following code on your custom task. See how to add a new task, for how to create a new task in MTEB.

from mteb import MTEB
from mteb.abstasks.AbsTaskReranking import AbsTaskReranking
from sentence_transformers import SentenceTransformer


class MyCustomTask(AbsTaskReranking):
    ...

model = SentenceTransformer("average_word_embeddings_komninos")
evaluation = MTEB(tasks=[MyCustomTask()])
evaluation.run(model)

Using a cross encoder for reranking

To use a cross encoder for reranking, you can directly use a CrossEncoder from SentenceTransformers. The following code shows a two-stage run with the second stage reading results saved from the first stage.

from mteb import MTEB
import mteb
from sentence_transformers import CrossEncoder, SentenceTransformer

cross_encoder = CrossEncoder("cross-encoder/ms-marco-TinyBERT-L-2-v2")
dual_encoder = SentenceTransformer("all-MiniLM-L6-v2")

tasks = mteb.get_tasks(tasks=["NFCorpus"], languages=["eng"])

subset = "default" # subset name used in the NFCorpus dataset
eval_splits = ["test"]

evaluation = MTEB(tasks=tasks)
evaluation.run(
    dual_encoder,
    eval_splits=eval_splits,
    save_predictions=True,
    output_folder="results/stage1",
)
evaluation.run(
    cross_encoder,
    eval_splits=eval_splits,
    top_k=5,
    save_predictions=True,
    output_folder="results/stage2",
    previous_results=f"results/stage1/NFCorpus_{subset}_predictions.json",
)

Using Late Interaction models for retrieval

from mteb import MTEB
import mteb


colbert = mteb.get_model("colbert-ir/colbertv2.0")
tasks = mteb.get_tasks(tasks=["NFCorpus"], languages=["eng"])

eval_splits = ["test"]

evaluation = MTEB(tasks=tasks)

evaluation.run(
    colbert,
    eval_splits=eval_splits,
    corpus_chunk_size=500,
)

This implementation employs the MaxSim operation to compute the similarity between sentences. While MaxSim provides high-quality results, it processes a larger number of embeddings, potentially leading to increased resource usage. To manage resource consumption, consider lowering the corpus_chunk_size parameter.

Saving retrieval task predictions

To save the predictions from a retrieval task, add the --save_predictions flag in the CLI or set save_predictions=True in the run method. The filename will be in the "{task_name}_{subset}_predictions.json" format.

Python:

from mteb import MTEB
import mteb
from sentence_transformers import SentenceTransformer

model = SentenceTransformer("all-MiniLM-L6-v2")

tasks = mteb.get_tasks(tasks=["NFCorpus"], languages=["eng"])

evaluation = MTEB(tasks=tasks)
evaluation.run(
    model,
    eval_splits=["test"],
    save_predictions=True,
    output_folder="results",
)

CLI:

mteb run -t NFCorpus -m all-MiniLM-L6-v2 --output_folder results --save_predictions

Fetching results from the results repository

Multiple models have already been run on tasks available within MTEB. These results are available results repository.

To make the results more easily accessible, we have designed custom functionality for retrieving from the repository. For instance, if you are selecting the best model for your French and English retrieval task on legal documents you could fetch the relevant tasks and create a dataframe of the results using the following code:

import mteb
from mteb.task_selection import results_to_dataframe

tasks = mteb.get_tasks(
    task_types=["Retrieval"], languages=["eng", "fra"], domains=["Legal"]
)

model_names = [
    "GritLM/GritLM-7B",
    "intfloat/multilingual-e5-small",
    "intfloat/multilingual-e5-base",
    "intfloat/multilingual-e5-large",
]
models = [mteb.get_model_meta(name) for name in model_names]

results = mteb.load_results(models=models, tasks=tasks)

df = results_to_dataframe(results)

Annotate Contamination

have your found contamination in the training data of a model? Please let us know, either by opening an issue or ideally by submitting a PR annotatig the training datasets of the model:

model_w_contamination = ModelMeta(
    name = "model-with-contamination"
    ...
    training_datasets: {"ArguAna": # name of dataset within MTEB
                        ["test"]} # the splits that have been trained on
    ...
)

Caching Embeddings To Re-Use Them

There are times you may want to cache the embeddings so you can re-use them. This may be true if you have multiple query sets for the same corpus (e.g. Wikipedia) or are doing some optimization over the queries (e.g. prompting, other experiments). You can setup a cache by using a simple wrapper, which will save the cache per task in the cache_embeddings/{task_name} folder:

# define your task and model above as normal
...
# wrap the model with the cache wrapper
from mteb.models.cache_wrapper import CachedEmbeddingWrapper
model_with_cached_emb = CachedEmbeddingWrapper(model, cache_path='path_to_cache_dir')
# run as normal
evaluation.run(model, ...) 

Documentation

Citing

MTEB was introduced in "MTEB: Massive Text Embedding Benchmark", feel free to cite:

@article{muennighoff2022mteb,
  doi = {10.48550/ARXIV.2210.07316},
  url = {https://arxiv.org/abs/2210.07316},
  author = {Muennighoff, Niklas and Tazi, Nouamane and Magne, Lo{\"\i}c and Reimers, Nils},
  title = {MTEB: Massive Text Embedding Benchmark},
  publisher = {arXiv},
  journal={arXiv preprint arXiv:2210.07316},  
  year = {2022}
}

You may also want to read and cite the amazing work that has extended MTEB & integrated new datasets:

• Shitao Xiao, Zheng Liu, Peitian Zhang, Niklas Muennighoff. "C-Pack: Packaged Resources To Advance General Chinese Embedding" arXiv 2023
• Michael Günther, Jackmin Ong, Isabelle Mohr, Alaeddine Abdessalem, Tanguy Abel, Mohammad Kalim Akram, Susana Guzman, Georgios Mastrapas, Saba Sturua, Bo Wang, Maximilian Werk, Nan Wang, Han Xiao. "Jina Embeddings 2: 8192-Token General-Purpose Text Embeddings for Long Documents" arXiv 2023
• Silvan Wehrli, Bert Arnrich, Christopher Irrgang. "German Text Embedding Clustering Benchmark" arXiv 2024
• Orion Weller, Benjamin Chang, Sean MacAvaney, Kyle Lo, Arman Cohan, Benjamin Van Durme, Dawn Lawrie, Luca Soldaini. "FollowIR: Evaluating and Teaching Information Retrieval Models to Follow Instructions" arXiv 2024
• Dawei Zhu, Liang Wang, Nan Yang, Yifan Song, Wenhao Wu, Furu Wei, Sujian Li. "LongEmbed: Extending Embedding Models for Long Context Retrieval" arXiv 2024
• Kenneth Enevoldsen, Márton Kardos, Niklas Muennighoff, Kristoffer Laigaard Nielbo. "The Scandinavian Embedding Benchmarks: Comprehensive Assessment of Multilingual and Monolingual Text Embedding" arXiv 2024

For works that have used MTEB for benchmarking, you can find them on the leaderboard.