from datasets import load_dataset
import gradio as gr
from huggingface_hub import HfApi, hf_hub_download
from huggingface_hub.repocard import metadata_load
import pandas as pd
TASKS = [
"BitextMining",
"Classification",
"Clustering",
"PairClassification",
"Reranking",
"Retrieval",
"STS",
"Summarization",
]
TASK_LIST_CLASSIFICATION = [
"AmazonCounterfactualClassification (en)",
"AmazonPolarityClassification",
"AmazonReviewsClassification (en)",
"Banking77Classification",
"EmotionClassification",
"ImdbClassification",
"MassiveIntentClassification (en)",
"MassiveScenarioClassification (en)",
"MTOPDomainClassification (en)",
"MTOPIntentClassification (en)",
"ToxicConversationsClassification",
"TweetSentimentExtractionClassification",
]
TASK_LIST_CLASSIFICATION_NORM = [x.replace(" (en)", "") for x in TASK_LIST_CLASSIFICATION]
TASK_LIST_CLUSTERING = [
"ArxivClusteringP2P",
"ArxivClusteringS2S",
"BiorxivClusteringP2P",
"BiorxivClusteringS2S",
"MedrxivClusteringP2P",
"MedrxivClusteringS2S",
"RedditClustering",
"RedditClusteringP2P",
"StackExchangeClustering",
"StackExchangeClusteringP2P",
"TwentyNewsgroupsClustering",
]
TASK_LIST_PAIR_CLASSIFICATION = [
"SprintDuplicateQuestions",
"TwitterSemEval2015",
"TwitterURLCorpus",
]
TASK_LIST_RERANKING = [
"AskUbuntuDupQuestions",
"MindSmallReranking",
"SciDocsRR",
"StackOverflowDupQuestions",
]
TASK_LIST_RETRIEVAL = [
"ArguAna",
"ClimateFEVER",
"CQADupstackRetrieval",
"DBPedia",
"FEVER",
"FiQA2018",
"HotpotQA",
"MSMARCO",
"NFCorpus",
"NQ",
"QuoraRetrieval",
"SCIDOCS",
"SciFact",
"Touche2020",
"TRECCOVID",
]
TASK_LIST_RETRIEVAL_NORM = TASK_LIST_RETRIEVAL + ["CQADupstackAndroidRetrieval",
"CQADupstackEnglishRetrieval",
"CQADupstackGamingRetrieval",
"CQADupstackGisRetrieval",
"CQADupstackMathematicaRetrieval",
"CQADupstackPhysicsRetrieval",
"CQADupstackProgrammersRetrieval",
"CQADupstackStatsRetrieval",
"CQADupstackTexRetrieval",
"CQADupstackUnixRetrieval",
"CQADupstackWebmastersRetrieval",
"CQADupstackWordpressRetrieval"
]
TASK_LIST_STS = [
"BIOSSES",
"SICK-R",
"STS12",
"STS13",
"STS14",
"STS15",
"STS16",
"STS17 (en-en)",
"STS22 (en)",
"STSBenchmark",
]
TASK_LIST_STS_NORM = [x.replace(" (en)", "").replace(" (en-en)", "") for x in TASK_LIST_STS]
TASK_LIST_SUMMARIZATION = [
"SummEval",
]
TASK_LIST_EN = TASK_LIST_CLASSIFICATION + TASK_LIST_CLUSTERING + TASK_LIST_PAIR_CLASSIFICATION + TASK_LIST_RERANKING + TASK_LIST_RETRIEVAL + TASK_LIST_STS + TASK_LIST_SUMMARIZATION
TASK_TO_METRIC = {
"BitextMining": "f1",
"Clustering": "v_measure",
"Classification": "accuracy",
"PairClassification": "cos_sim_ap",
"Reranking": "map",
"Retrieval": "ndcg_at_10",
"STS": "cos_sim_spearman",
"Summarization": "cos_sim_spearman",
}
def make_clickable_model(model_name, link=None):
if link is None:
link = "https://huggingface.co/" + model_name
# Remove user from model name
return (
f'{model_name.split("/")[-1]}'
)
# Models without metadata, thus we cannot fetch their results naturally
EXTERNAL_MODELS = [
"LASER2",
"LaBSE",
"all-MiniLM-L12-v2",
"all-MiniLM-L6-v2",
"all-mpnet-base-v2",
"allenai-specter",
"bert-base-uncased",
"contriever-base-msmarco",
"glove.6B.300d",
"gtr-t5-base",
"gtr-t5-large",
"gtr-t5-xl",
"gtr-t5-xxl",
"komninos",
"msmarco-bert-co-condensor",
"paraphrase-multilingual-MiniLM-L12-v2",
"paraphrase-multilingual-mpnet-base-v2",
"sentence-t5-base",
"sentence-t5-large",
"sentence-t5-xl",
"sentence-t5-xxl",
"sgpt-bloom-1b3-nli",
"sgpt-bloom-7b1-msmarco",
"sup-simcse-bert-base-uncased",
"text-similarity-ada-001",
"unsup-simcse-bert-base-uncased",
]
EXTERNAL_MODEL_TO_LINK = {
"LASER2": "https://github.com/facebookresearch/LASER",
"text-similarity-ada-001": "https://beta.openai.com/docs/guides/embeddings/types-of-embedding-models",
}
EXTERNAL_MODEL_RESULTS = {model: {k: {v: []} for k, v in TASK_TO_METRIC.items()} for model in EXTERNAL_MODELS}
def add_lang(examples):
if not(examples["eval_language"]):
examples["mteb_dataset_name_with_lang"] = examples["mteb_dataset_name"]
else:
examples["mteb_dataset_name_with_lang"] = examples["mteb_dataset_name"] + f' ({examples["eval_language"]})'
return examples
def add_task(examples):
# Could be added to the dataset loading script instead
if examples["mteb_dataset_name"] in TASK_LIST_CLASSIFICATION_NORM:
examples["mteb_task"] = "Classification"
elif examples["mteb_dataset_name"] in TASK_LIST_CLUSTERING:
examples["mteb_task"] = "Clustering"
elif examples["mteb_dataset_name"] in TASK_LIST_PAIR_CLASSIFICATION:
examples["mteb_task"] = "PairClassification"
elif examples["mteb_dataset_name"] in TASK_LIST_RERANKING:
examples["mteb_task"] = "Reranking"
elif examples["mteb_dataset_name"] in TASK_LIST_RETRIEVAL_NORM:
examples["mteb_task"] = "Retrieval"
elif examples["mteb_dataset_name"] in TASK_LIST_STS_NORM:
examples["mteb_task"] = "STS"
elif examples["mteb_dataset_name"] in TASK_LIST_SUMMARIZATION:
examples["mteb_task"] = "Summarization"
else:
examples["mteb_task"] = "BitextMining"
return examples
for model in EXTERNAL_MODELS:
ds = load_dataset("mteb/results", model)
ds = ds.map(add_lang)
ds = ds.map(add_task)
base_dict = {"Model": make_clickable_model(model, link=EXTERNAL_MODEL_TO_LINK.get(model, "https://huggingface.co/spaces/mteb/leaderboard"))}
# For now only one metric per task - Could add more metrics lateron
for task, metric in TASK_TO_METRIC.items():
ds_dict = ds.filter(lambda x: (x["mteb_task"] == task) and (x["metric"] == metric))["test"].to_dict()
ds_dict = {k: round(v, 2) for k, v in zip(ds_dict["mteb_dataset_name_with_lang"], ds_dict["score"])}
EXTERNAL_MODEL_RESULTS[model][task][metric].append({**base_dict, **ds_dict})
def get_mteb_data(tasks=["Clustering"], langs=[], task_to_metric=TASK_TO_METRIC):
api = HfApi()
models = api.list_models(filter="mteb")
# Initialize list to models that we cannot fetch metadata from
df_list = []
for model in EXTERNAL_MODEL_RESULTS:
results_list = [res for task in tasks for res in EXTERNAL_MODEL_RESULTS[model][task][task_to_metric[task]]]
if langs:
# Would be cleaner to rely on an extra language column instead
langs_format = [f"({lang})" for lang in langs]
res = {k: v for d in results_list for k, v in d.items() if any([k.split(" ")[-1] in (k, x) for x in langs_format])}
else:
res = {k: v for d in results_list for k, v in d.items()}
# Model & at least one result
if len(res) > 1:
df_list.append(res)
for model in models:
readme_path = hf_hub_download(model.modelId, filename="README.md")
meta = metadata_load(readme_path)
# meta['model-index'][0]["results"] is list of elements like:
# {
# "task": {"type": "Classification"},
# "dataset": {
# "type": "mteb/amazon_massive_intent",
# "name": "MTEB MassiveIntentClassification (nb)",
# "config": "nb",
# "split": "test",
# },
# "metrics": [
# {"type": "accuracy", "value": 39.81506388702084},
# {"type": "f1", "value": 38.809586587791664},
# ],
# },
# Use "get" instead of dict indexing to skip incompat metadata instead of erroring out
if langs:
task_results = [sub_res for sub_res in meta["model-index"][0]["results"] if (sub_res.get("task", {}).get("type", "") in tasks) and (sub_res.get("dataset", {}).get("config", "default") in ("default", *langs))]
else:
task_results = [sub_res for sub_res in meta["model-index"][0]["results"] if (sub_res.get("task", {}).get("type", "") in tasks)]
out = [{res["dataset"]["name"].replace("MTEB ", ""): [round(score["value"], 2) for score in res["metrics"] if score["type"] == task_to_metric.get(res["task"]["type"])][0]} for res in task_results]
out = {k: v for d in out for k, v in d.items()}
out["Model"] = make_clickable_model(model.modelId)
df_list.append(out)
df = pd.DataFrame(df_list)
# Put 'Model' column first
cols = sorted(list(df.columns))
cols.insert(0, cols.pop(cols.index("Model")))
df = df[cols]
df.fillna("", inplace=True)
return df
def get_mteb_average():
global DATA_OVERALL, DATA_CLASSIFICATION_EN, DATA_CLUSTERING, DATA_PAIR_CLASSIFICATION, DATA_RERANKING, DATA_RETRIEVAL, DATA_STS_EN, DATA_SUMMARIZATION, NUM_SCORES
DATA_OVERALL = get_mteb_data(
tasks=[
"Classification",
"Clustering",
"PairClassification",
"Reranking",
"Retrieval",
"STS",
"Summarization",
],
langs=["en", "en-en"],
)
# Approximation (Missing Bitext Mining & including some nans)
NUM_SCORES = DATA_OVERALL.shape[0] * DATA_OVERALL.shape[1]
# Debugging:
# DATA_OVERALL.to_csv("overall.csv")
DATA_OVERALL.insert(1, f"Average ({len(TASK_LIST_EN)} datasets)", DATA_OVERALL[TASK_LIST_EN].mean(axis=1, skipna=False))
DATA_OVERALL.insert(2, f"Classification Average ({len(TASK_LIST_CLASSIFICATION)} datasets)", DATA_OVERALL[TASK_LIST_CLASSIFICATION].mean(axis=1, skipna=False))
DATA_OVERALL.insert(3, f"Clustering Average ({len(TASK_LIST_CLUSTERING)} datasets)", DATA_OVERALL[TASK_LIST_CLUSTERING].mean(axis=1, skipna=False))
DATA_OVERALL.insert(4, f"Pair Classification Average ({len(TASK_LIST_PAIR_CLASSIFICATION)} datasets)", DATA_OVERALL[TASK_LIST_PAIR_CLASSIFICATION].mean(axis=1, skipna=False))
DATA_OVERALL.insert(5, f"Reranking Average ({len(TASK_LIST_RERANKING)} datasets)", DATA_OVERALL[TASK_LIST_RERANKING].mean(axis=1, skipna=False))
DATA_OVERALL.insert(6, f"Retrieval Average ({len(TASK_LIST_RETRIEVAL)} datasets)", DATA_OVERALL[TASK_LIST_RETRIEVAL].mean(axis=1, skipna=False))
DATA_OVERALL.insert(7, f"STS Average ({len(TASK_LIST_STS)} datasets)", DATA_OVERALL[TASK_LIST_STS].mean(axis=1, skipna=False))
DATA_OVERALL.insert(8, f"Summarization Average ({len(TASK_LIST_SUMMARIZATION)} dataset)", DATA_OVERALL[TASK_LIST_SUMMARIZATION].mean(axis=1, skipna=False))
DATA_OVERALL.sort_values(f"Average ({len(TASK_LIST_EN)} datasets)", ascending=False, inplace=True)
# Start ranking from 1
DATA_OVERALL.insert(0, "Rank", list(range(1, len(DATA_OVERALL) + 1)))
DATA_OVERALL = DATA_OVERALL.round(2)
DATA_CLASSIFICATION_EN = DATA_OVERALL[["Model"] + TASK_LIST_CLASSIFICATION]
DATA_CLUSTERING = DATA_OVERALL[["Model"] + TASK_LIST_CLUSTERING]
DATA_PAIR_CLASSIFICATION = DATA_OVERALL[["Model"] + TASK_LIST_PAIR_CLASSIFICATION]
DATA_RERANKING = DATA_OVERALL[["Model"] + TASK_LIST_RERANKING]
DATA_RETRIEVAL = DATA_OVERALL[["Model"] + TASK_LIST_RETRIEVAL]
DATA_STS_EN = DATA_OVERALL[["Model"] + TASK_LIST_STS]
DATA_SUMMARIZATION = DATA_OVERALL[["Model"] + TASK_LIST_SUMMARIZATION]
DATA_OVERALL = DATA_OVERALL[["Rank", "Model", f"Average ({len(TASK_LIST_EN)} datasets)", f"Classification Average ({len(TASK_LIST_CLASSIFICATION)} datasets)", f"Clustering Average ({len(TASK_LIST_CLUSTERING)} datasets)", f"Pair Classification Average ({len(TASK_LIST_PAIR_CLASSIFICATION)} datasets)", f"Reranking Average ({len(TASK_LIST_RERANKING)} datasets)", f"Retrieval Average ({len(TASK_LIST_RETRIEVAL)} datasets)", f"STS Average ({len(TASK_LIST_STS)} datasets)", f"Summarization Average ({len(TASK_LIST_SUMMARIZATION)} dataset)"]]
return DATA_OVERALL
get_mteb_average()
block = gr.Blocks()
with block:
gr.Markdown(f"""
Massive Text Embedding Benchmark (MTEB) Leaderboard. To submit, refer to the MTEB GitHub repository 🤗
- **Total Datasets**: 56
- **Total Languages**: 117
- **Total Scores**: >{NUM_SCORES}
- **Total Models**: {len(DATA_OVERALL)}
""")
with gr.Tabs():
with gr.TabItem("Overall"):
with gr.Row():
gr.Markdown("""
**Overall MTEB English leaderboard 🔮**
- **Metric:** Various, refer to task tabs
- **Languages:** English, refer to task tabs for others
""")
with gr.Row():
data_overall = gr.components.Dataframe(
DATA_OVERALL,
datatype=["number", "markdown"] + ["number"] * len(DATA_OVERALL.columns),
type="pandas",
wrap=True,
)
with gr.Row():
data_run = gr.Button("Refresh")
data_run.click(get_mteb_average, inputs=None, outputs=data_overall)
with gr.TabItem("BitextMining"):
with gr.Row():
gr.Markdown("""
**Bitext Mining Leaderboard 🎌**
- **Metric:** F1 (f1)
- **Languages:** 117
""")
with gr.Row():
data_bitext_mining = gr.components.Dataframe(
datatype=["markdown"] + ["number"] * 500, # hack when we don't know how many columns
type="pandas",
)
with gr.Row():
data_run = gr.Button("Refresh")
task_bitext_mining = gr.Variable(value=["BitextMining"])
data_run.click(
get_mteb_data,
inputs=[task_bitext_mining],
outputs=data_bitext_mining,
)
with gr.TabItem("Classification"):
with gr.TabItem("English"):
with gr.Row():
gr.Markdown("""
**Classification Leaderboard ❤️**
- **Metric:** Accuracy (accuracy)
- **Languages:** English
""")
with gr.Row():
data_classification_en = gr.components.Dataframe(
DATA_CLASSIFICATION_EN,
datatype=["markdown"] + ["number"] * len(DATA_CLASSIFICATION_EN.columns),
type="pandas",
)
with gr.Row():
data_run_classification_en = gr.Button("Refresh")
task_classification_en = gr.Variable(value=["Classification"])
lang_classification_en = gr.Variable(value=["en"])
data_run_classification_en.click(
get_mteb_data,
inputs=[
task_classification_en,
lang_classification_en,
],
outputs=data_classification_en,
)
with gr.TabItem("Multilingual"):
with gr.Row():
gr.Markdown("""
**Classification Multilingual Leaderboard 💜💚💙**
- **Metric:** Accuracy (accuracy)
- **Languages:** 51
""")
with gr.Row():
data_classification = gr.components.Dataframe(
datatype=["markdown"] + ["number"] * 200, # hack when we don't know how many columns
type="pandas",
)
with gr.Row():
data_run = gr.Button("Refresh")
task_classification = gr.Variable(value=["Classification"])
data_run.click(
get_mteb_data,
inputs=[task_classification],
outputs=data_classification,
)
with gr.TabItem("Clustering"):
with gr.Row():
gr.Markdown("""
**Clustering Leaderboard ✨**
- **Metric:** Validity Measure (v_measure)
- **Languages:** English
""")
with gr.Row():
data_clustering = gr.components.Dataframe(
DATA_CLUSTERING,
datatype=["markdown"] + ["number"] * len(DATA_CLUSTERING.columns),
type="pandas",
)
with gr.Row():
data_run = gr.Button("Refresh")
task_clustering = gr.Variable(value=["Clustering"])
data_run.click(
get_mteb_data,
inputs=[task_clustering],
outputs=data_clustering,
)
with gr.TabItem("Pair Classification"):
with gr.Row():
gr.Markdown("""
**Pair Classification Leaderboard 🎭**
- **Metric:** Average Precision based on Cosine Similarities (cos_sim_ap)
- **Languages:** English
""")
with gr.Row():
data_pair_classification = gr.components.Dataframe(
DATA_PAIR_CLASSIFICATION,
datatype=["markdown"] + ["number"] * len(DATA_PAIR_CLASSIFICATION.columns),
type="pandas",
)
with gr.Row():
data_run = gr.Button("Refresh")
task_pair_classification = gr.Variable(value=["PairClassification"])
data_run.click(
get_mteb_data,
inputs=[task_pair_classification],
outputs=data_pair_classification,
)
with gr.TabItem("Retrieval"):
with gr.Row():
gr.Markdown("""
**Retrieval Leaderboard 🔎**
- **Metric:** Normalized Discounted Cumulative Gain @ k (ndcg_at_10)
- **Languages:** English
""")
with gr.Row():
data_retrieval = gr.components.Dataframe(
DATA_RETRIEVAL,
# Add support for more columns than existing as a buffer for CQADupstack & other Retrieval tasks (e.g. MSMARCOv2)
datatype=["markdown"] + ["number"] * len(DATA_RETRIEVAL.columns) * 2,
type="pandas",
)
with gr.Row():
data_run = gr.Button("Refresh")
task_retrieval = gr.Variable(value=["Retrieval"])
data_run.click(
get_mteb_data, inputs=[task_retrieval], outputs=data_retrieval
)
with gr.TabItem("Reranking"):
with gr.Row():
gr.Markdown("""
**Reranking Leaderboard 🥈**
- **Metric:** Mean Average Precision (MAP)
- **Languages:** English
""")
with gr.Row():
data_reranking = gr.components.Dataframe(
DATA_RERANKING,
datatype=["markdown"] + ["number"] * len(DATA_RERANKING.columns),
type="pandas",
)
with gr.Row():
data_run = gr.Button("Refresh")
task_reranking = gr.Variable(value=["Reranking"])
metric_reranking = gr.Variable(value="map")
data_run.click(
get_mteb_data, inputs=[task_reranking], outputs=data_reranking
)
with gr.TabItem("STS"):
with gr.TabItem("English"):
with gr.Row():
gr.Markdown("""
**STS Leaderboard 🤖**
- **Metric:** Spearman correlation based on cosine similarity
- **Languages:** English
""")
with gr.Row():
data_sts_en = gr.components.Dataframe(
DATA_STS_EN,
datatype=["markdown"] + ["number"] * len(DATA_STS_EN.columns),
type="pandas",
)
with gr.Row():
data_run_sts_en = gr.Button("Refresh")
task_sts_en = gr.Variable(value=["STS"])
lang_sts_en = gr.Variable(value=["en", "en-en"])
data_run_sts_en.click(
get_mteb_data,
inputs=[task_sts_en, lang_sts_en],
outputs=data_sts_en,
)
with gr.TabItem("Multilingual"):
with gr.Row():
gr.Markdown("""
**STS Multilingual Leaderboard 👽**
- **Metric:** Spearman correlation based on cosine similarity
- **Languages:** Arabic, Chinese, Dutch, English, French, German, Italian, Korean, Polish, Russian, Spanish
""")
with gr.Row():
data_sts = gr.components.Dataframe(
datatype=["markdown"] + ["number"] * 100, # hack when we don't know how many columns
type="pandas",
)
with gr.Row():
data_run = gr.Button("Refresh")
task_sts = gr.Variable(value=["STS"])
data_run.click(get_mteb_data, inputs=[task_sts], outputs=data_sts)
with gr.TabItem("Summarization"):
with gr.Row():
gr.Markdown("""
**Summarization Leaderboard 📜**
- **Metric:** Spearman correlation based on cosine similarity
- **Languages:** English
""")
with gr.Row():
data_summarization = gr.components.Dataframe(
DATA_SUMMARIZATION,
datatype=["markdown"] + ["number"] * 2,
type="pandas",
)
with gr.Row():
data_run = gr.Button("Refresh")
task_summarization = gr.Variable(value=["Summarization"])
data_run.click(
get_mteb_data,
inputs=[task_summarization],
outputs=data_summarization,
)
gr.Markdown(f"""
Made with ❤️ for NLP by Niklas Muennighoff.
""")
# Running the function on page load in addition to when the button is clicked
# This is optional - If deactivated the data created loaded at "Build time" is shown like for Overall tab
block.load(get_mteb_data, inputs=[task_bitext_mining], outputs=data_bitext_mining)
block.load(get_mteb_data, inputs=[task_classification_en, lang_classification_en], outputs=data_classification_en)
block.load(get_mteb_data, inputs=[task_classification], outputs=data_classification)
block.load(get_mteb_data, inputs=[task_clustering], outputs=data_clustering)
block.load(get_mteb_data, inputs=[task_pair_classification], outputs=data_pair_classification)
block.load(get_mteb_data, inputs=[task_retrieval], outputs=data_retrieval)
block.load(get_mteb_data, inputs=[task_reranking], outputs=data_reranking)
block.load(get_mteb_data, inputs=[task_sts_en, lang_sts_en], outputs=data_sts_en)
block.load(get_mteb_data, inputs=[task_sts], outputs=data_sts)
block.load(get_mteb_data, inputs=[task_summarization], outputs=data_summarization)
block.launch()
# Possible changes:
# Could check if tasks are valid (Currently users could just invent new tasks - similar for languages)
# Could make it load in the background without the Gradio logo closer to the Deep RL space
# Could add graphs / other visual content
# Could add verification marks
# Sources:
# https://huggingface.co/spaces/gradio/leaderboard
# https://huggingface.co/spaces/huggingface-projects/Deep-Reinforcement-Learning-Leaderboard
# https://getemoji.com/