fix: resolve circular imports and add CI enforcement

mteb/abstasks/task_metadata.py

@@ -20,7 +20,6 @@
 )
 from typing_extensions import Required, TypedDict  # noqa: TC002
 
-import mteb
 from mteb.languages import check_language_code
 from mteb.types import (
     Languages,
@@ -588,9 +587,10 @@ def _create_dataset_card_data(
             multilinguality = "translated"
 
         if self.adapted_from is not None:
+            from mteb.get_tasks import get_tasks
+
             source_datasets = [
-                task.metadata.dataset["path"]
-                for task in mteb.get_tasks(self.adapted_from)
+                task.metadata.dataset["path"] for task in get_tasks(self.adapted_from)
             ]
             source_datasets.append(self.dataset["path"])
         else:

mteb/benchmarks/_create_table.py

@@ -8,8 +8,8 @@
 import numpy as np
 import pandas as pd
 
-import mteb
 from mteb.get_tasks import get_task, get_tasks
+from mteb.models.get_model_meta import get_model_meta
 
 if TYPE_CHECKING:
     from mteb.results.benchmark_results import BenchmarkResults
@@ -145,7 +145,7 @@ def _create_summary_table_from_benchmark_results(
     joint_table = joint_table.reset_index()
 
     # Add model metadata
-    model_metas = joint_table["model_name"].map(mteb.get_model_meta)
+    model_metas = joint_table["model_name"].map(get_model_meta)
     joint_table = joint_table[model_metas.notna()]
     joint_table["model_link"] = model_metas.map(lambda m: m.reference)
 
@@ -386,7 +386,7 @@ def _create_summary_table_mean_public_private(
     joint_table = joint_table.reset_index()
 
     # Add model metadata
-    model_metas = joint_table["model_name"].map(mteb.get_model_meta)
+    model_metas = joint_table["model_name"].map(get_model_meta)
     joint_table = joint_table[model_metas.notna()]
     joint_table["model_link"] = model_metas.map(lambda m: m.reference)
 
@@ -505,7 +505,7 @@ def _create_summary_table_mean_subset(
     joint_table = joint_table.reset_index()
 
     # Add model metadata
-    model_metas = joint_table["model_name"].map(mteb.get_model_meta)
+    model_metas = joint_table["model_name"].map(get_model_meta)
     joint_table = joint_table[model_metas.notna()]
     joint_table["model_link"] = model_metas.map(lambda m: m.reference)
 
@@ -622,7 +622,7 @@ def _create_summary_table_mean_task_type(
     joint_table = joint_table.reset_index()
 
     # Add model metadata
-    model_metas = joint_table["model_name"].map(mteb.get_model_meta)
+    model_metas = joint_table["model_name"].map(get_model_meta)
     joint_table = joint_table[model_metas.notna()]
     joint_table["model_link"] = model_metas.map(lambda m: m.reference)
 

mteb/cache.py

@@ -16,11 +16,12 @@
 import requests
 from pydantic import ValidationError
 
-import mteb
 from mteb._helpful_enum import HelpfulStrEnum
 from mteb.abstasks import AbsTask
 from mteb.benchmarks.benchmark import Benchmark
+from mteb.benchmarks.get_benchmark import get_benchmark
 from mteb.models import ModelMeta
+from mteb.models.get_model_meta import get_model_metas
 from mteb.models.model_meta import _serialize_experiment_kwargs_to_name
 from mteb.results import BenchmarkResults, ModelResult, TaskResult
 
@@ -548,7 +549,7 @@ def _rebuild_from_full_repository(self, quick_cache_path: Path) -> BenchmarkResu
 
         all_model_names = [
             model_meta.name
-            for model_meta in mteb.get_model_metas()
+            for model_meta in get_model_metas()
             if model_meta.name is not None
         ]
 
@@ -861,7 +862,7 @@ def load_results(
             ... )
         """
         if isinstance(tasks, str):
-            tasks = mteb.get_benchmark(tasks)
+            tasks = get_benchmark(tasks)
 
         if isinstance(load_experiments, str):
             load_experiments = LoadExperimentEnum.from_str(load_experiments)

mteb/deprecated_evaluator.py

@@ -15,9 +15,9 @@
 
 import datasets
 
-import mteb
 from mteb.abstasks import AbsTask
 from mteb.benchmarks import Benchmark
+from mteb.get_tasks import get_tasks
 from mteb.models import (
     CrossEncoderWrapper,
     ModelMeta,
@@ -156,7 +156,7 @@ def mteb_benchmarks(self):
     @classmethod
     def mteb_tasks(cls):
         """Get all tasks available in the MTEB."""
-        tasks = mteb.get_tasks()
+        tasks = get_tasks()
         instance = cls(tasks)
         instance._display_tasks(tasks, name="MTEB tasks")
 

mteb/models/abs_encoder.py

@@ -5,7 +5,6 @@
 from abc import ABC, abstractmethod
 from typing import TYPE_CHECKING, Any, Literal, cast, get_args, overload
 
-import mteb
 from mteb.abstasks.task_metadata import TaskType
 from mteb.similarity_functions import (
     cos_sim,
@@ -166,7 +165,9 @@ def validate_task_to_prompt_name(
 
             if task_name not in task_types and task_name not in prompt_types:
                 try:
-                    mteb.get_task(task_name=task_name)
+                    from mteb.get_tasks import get_task
+
+                    get_task(task_name=task_name)
                 except KeyError:
                     msg = f"Task name {task_name} is not valid. {valid_keys_msg}"
                     logger.warning(msg)
@@ -224,7 +225,9 @@ def get_instruction(
         if prompt:
             return prompt
 
-        abstask = mteb.get_task(task_name=task_metadata.name)
+        from mteb.get_tasks import get_task
+
+        abstask = get_task(task_name=task_metadata.name)
         return abstask.abstask_prompt
 
     def format_instruction(

mteb/results/task_result.py

@@ -15,8 +15,6 @@
 from pydantic import BaseModel, field_validator
 from typing_extensions import deprecated
 
-import mteb
-from mteb import TaskMetadata
 from mteb._helpful_enum import HelpfulStrEnum
 from mteb._hf_integration.eval_result_model import (
     HFEvalResult,
@@ -26,6 +24,7 @@
 )
 from mteb.abstasks import AbsTaskClassification
 from mteb.abstasks.abstask import AbsTask
+from mteb.abstasks.task_metadata import TaskMetadata
 from mteb.languages import LanguageScripts
 from mteb.models.model_meta import ScoringFunction
 from mteb.types import (
@@ -909,7 +908,9 @@ def get_hf_eval_results(self) -> list[EvalResult]:
         return results
 
     def _to_hf_benchmark_result(self, user: str | None = None) -> HFEvalResults:
-        task_metadata = mteb.get_task(self.task_name).metadata
+        from mteb.get_tasks import get_task
+
+        task_metadata = get_task(self.task_name).metadata
         dataset_id = task_metadata.dataset["path"]
         dataset_revision = task_metadata.dataset["revision"]
         eval_results = []

tests/test_result_cache_load_from_cache.py

@@ -98,7 +98,7 @@ def test_full_rebuild_process(self, tmp_path):
         with (
             patch.object(cache, "download_from_remote") as mock_download,
             patch.object(cache, "load_results") as mock_load_results,
-            patch("mteb.get_model_metas") as mock_get_model_metas,
+            patch("mteb.cache.get_model_metas") as mock_get_model_metas,
         ):
             # Mock model metas - None names should be filtered
             meta1 = MagicMock()
@@ -138,7 +138,7 @@ def test_rebuild_error_propagation(self, tmp_path):
         with (
             patch.object(cache, "download_from_remote"),
             patch.object(cache, "load_results") as mock_load_results,
-            patch("mteb.get_model_metas") as mock_get_model_metas,
+            patch("mteb.cache.get_model_metas") as mock_get_model_metas,
         ):
             meta = MagicMock()
             meta.name = "model1"