mteb 2.5.2__py3-none-any.whl → 2.5.4__py3-none-any.whl

mteb/_create_dataloaders.py

@@ -1,4 +1,5 @@
 import logging
+import warnings
 from collections.abc import Callable
 from typing import Any, cast
 
@@ -113,11 +114,8 @@ def _create_text_dataloader_for_queries(
     )
 
 
-_warned_about_user_role = False
-
-
 def _convert_conv_history_to_query(
-    row: dict[str, list[str] | Conversation],
+    row: dict[str, str | list[str] | Conversation],
 ) -> dict[str, str | Conversation]:
     """Convert a conversation history to a single query string.
 
@@ -127,21 +125,18 @@ def _convert_conv_history_to_query(
     Returns:
         The updated row with the "query" and "text" fields set to the conversation string, and the "conversation" field set to the list of ConversationTurn.
     """
-    global _warned_about_user_role
-
     conversation = row["text"]
     # if it's a list of strings, just join them
     if isinstance(conversation, list) and isinstance(conversation[0], str):
-        conversation = cast(list[str], conversation)
-        conv_str = "; ".join(conversation)
+        conversation_ = cast(list[str], conversation)
+        conv_str = "; ".join(conversation_)
         current_conversation = [
-            ConversationTurn(role="user", content=message) for message in conversation
+            ConversationTurn(role="user", content=message) for message in conversation_
         ]
-        if not _warned_about_user_role:
-            logger.warning(
-                "Conversations are a list of strings. Used 'user' role for all turns."
-            )
-            _warned_about_user_role = True
+        warnings.warn(
+            "Conversations are a list of strings. Used 'user' role for all turns.",
+            category=UserWarning,
+        )
     # otherwise, it's a list of dictionaries, which we need to convert to strings
     elif isinstance(conversation, list) and isinstance(conversation[0], dict):
         conv = []
@@ -178,7 +173,7 @@ def _convert_conv_history_to_query(
 
     row["text"] = conv_str
     row["conversation"] = current_conversation
-    return row
+    return cast(dict[str, str | list[ConversationTurn]], row)
 
 
 def _create_dataloader_for_queries_conversation(

mteb/_evaluators/any_sts_evaluator.py

@@ -57,10 +57,7 @@ class AnySTSEvaluator(Evaluator):
         self.input2_prompt_type = input2_prompt_type
 
     def __call__(
-        self,
-        model: EncoderProtocol,
-        *,
-        encode_kwargs: dict[str, Any],
+        self, model: EncoderProtocol, *, encode_kwargs: dict[str, Any]
     ) -> STSEvaluatorScores:
         logger.info("Running semantic similarity - Encoding samples (1/2)")
         embeddings1 = model.encode(

mteb/_evaluators/evaluator.py

@@ -1,4 +1,5 @@
 from abc import ABC, abstractmethod
+from collections.abc import Iterable, Mapping
 from typing import Any
 
 from mteb.abstasks.abstask import _set_seed
@@ -18,7 +19,7 @@ class Evaluator(ABC):
     @abstractmethod
     def __call__(
         self, model: EncoderProtocol, *, encode_kwargs: dict[str, Any]
-    ) -> dict[str, float]:
+    ) -> Mapping[str, float] | Iterable[Any]:
         """This is called during training to evaluate the model.
 
         It returns scores.

mteb/_evaluators/image/imagetext_pairclassification_evaluator.py

@@ -1,6 +1,7 @@
 from __future__ import annotations
 
 import logging
+from collections.abc import Sequence
 from typing import TYPE_CHECKING, Any
 
 import torch
@@ -61,8 +62,8 @@ class ImageTextPairClassificationEvaluator(Evaluator):
     def __init__(
         self,
         dataset,
-        images_column_names: str | list[str],
-        texts_column_names: str | list[str],
+        images_column_names: str | Sequence[str],
+        texts_column_names: str | Sequence[str],
         num_images_per_sample: int,
         num_texts_per_sample: int,
         task_metadata: TaskMetadata,
@@ -82,10 +83,8 @@ class ImageTextPairClassificationEvaluator(Evaluator):
         self.hf_split = hf_split
         self.hf_subset = hf_subset
 
-    def __call__(
-        self,
-        model: EncoderProtocol,
-        encode_kwargs: dict[str, Any],
+    def __call__(  # type: ignore[override]
+        self, model: EncoderProtocol, *, encode_kwargs: dict[str, Any]
     ) -> list[torch.Tensor]:
         images = []
         if isinstance(self.images_column_names, str):

mteb/_evaluators/pair_classification_evaluator.py

@@ -148,7 +148,9 @@ class PairClassificationEvaluator(Evaluator):
         hf_subset: str,
         **encode_kwargs: Any,
     ) -> np.ndarray:
-        index_map, all_unique_texts, all_texts_indexes = {}, [], []
+        index_map = {}
+        all_unique_texts: list[str] = []
+        all_texts_indexes = []
         for text in all_texts:
             text_hash = hash(text)
             if text_hash not in index_map:

mteb/_evaluators/retrieval_metrics.py

@@ -1,5 +1,6 @@
 import logging
 from collections import defaultdict
+from collections.abc import Mapping
 from typing import Any
 
 import numpy as np
@@ -15,7 +16,7 @@ logger = logging.getLogger(__name__)
 
 def mrr(
     qrels: RelevantDocumentsType,
-    results: dict[str, dict[str, float]],
+    results: Mapping[str, Mapping[str, float]],
     k_values: list[int],
 ) -> dict[str, list[float]]:
     mrr_metrics = defaultdict(list)
@@ -32,7 +33,7 @@ def mrr(
             doc_id for doc_id in qrels[query_id] if qrels[query_id][doc_id] > 0
         }
         for k in k_values:
-            rr = 0
+            rr = 0.0
             for rank, hit in enumerate(top_hits[query_id][0:k]):
                 if hit[0] in query_relevant_docs:
                     rr = 1.0 / (rank + 1)
@@ -45,8 +46,8 @@ def recall_cap(
     qrels: RelevantDocumentsType,
     results: dict[str, dict[str, float]],
     k_values: list[int],
-) -> dict[str, list[float]]:
-    capped_recall = defaultdict(list)
+) -> dict[str, list[float | None]]:
+    capped_recall: dict[str, list[float | None]] = defaultdict(list)
 
     k_max = max(k_values)
 
@@ -188,7 +189,7 @@ def evaluate_p_mrr_change(
     Returns:
         A dictionary with the scores, including "p-MRR", "og" and "changed" keys.
     """
-    followir_scores = defaultdict(dict)
+    followir_scores: dict[str, float | dict[str, float]] = defaultdict(dict)
 
     qrels_sep = {
         "og": {k: v for k, v in qrels.items() if k.endswith("-og")},
@@ -227,7 +228,7 @@ def evaluate_p_mrr_change(
             ndcg, _map, recall, precision, naucs, avg_mrr, naucs_mrr, cv_recall, {}
         )
         for key, value in scores_dict.items():
-            followir_scores[name][key] = value
+            followir_scores[name][key] = value  # type: ignore[index]
 
     return followir_scores
 
@@ -254,8 +255,8 @@ def confidence_scores(sim_scores: list[float]) -> dict[str, float]:
     sim_scores_sorted = sorted(sim_scores)[::-1]
 
     cs_max = sim_scores_sorted[0]
-    cs_std = np.std(sim_scores)
-    cs_diff1 = None
+    cs_std = float(np.std(sim_scores))
+    cs_diff1 = 0.0
     if len(sim_scores) > 1:
         cs_diff1 = sim_scores_sorted[0] - sim_scores_sorted[1]
     elif len(sim_scores) == 1:
@@ -410,7 +411,7 @@ def make_score_dict(
     cv_recall: dict[str, float],
     task_scores: dict[str, float],
     previous_results_model_meta: dict[str, Any] | None = None,
-) -> dict[str, float]:
+) -> dict[str, Any]:
     return {
         **{f"ndcg_at_{k.split('@')[1]}": v for (k, v) in ndcg.items()},
         **{f"map_at_{k.split('@')[1]}": v for (k, v) in _map.items()},
@@ -528,7 +529,7 @@ def max_over_subqueries(
 
 
 def calculate_retrieval_scores(
-    results: dict[str, dict[str, float]],
+    results: Mapping[str, Mapping[str, float]],
     qrels: RelevantDocumentsType,
     k_values: list[int],
     skip_first_result: bool = False,
@@ -576,7 +577,7 @@ def calculate_retrieval_scores(
 
 
 def evaluate_abstention(
-    results: dict[str, dict[str, float]],
+    results: Mapping[str, Mapping[str, float]],
     metric_scores: dict[str, list[float]],
 ) -> dict[str, float]:
     """Computes normalized Area Under the Curve on a set of evaluated instances as presented in the paper https://arxiv.org/abs/2402.12997
@@ -591,21 +592,21 @@ def evaluate_abstention(
     all_sim_scores = [list(results[qid].values()) for qid in list(results.keys())]
     all_conf_scores = [confidence_scores(sim_scores) for sim_scores in all_sim_scores]
     conf_fcts = list(all_conf_scores[0].keys())
-    all_conf_scores = {
+    all_conf_scores_ = {
         fct: np.array([x[fct] for x in all_conf_scores]) for fct in conf_fcts
     }
-    metric_scores = {k: np.array(v) for k, v in metric_scores.items()}
+    metric_scores_ = {k: np.array(v) for k, v in metric_scores.items()}
     naucs = {}
 
-    for metric_name, scores in metric_scores.items():
-        for fct, conf_scores in all_conf_scores.items():
+    for metric_name, scores in metric_scores_.items():
+        for fct, conf_scores in all_conf_scores_.items():
             naucs[f"nAUC_{metric_name}_{fct}"] = nauc(conf_scores, scores)
 
     return naucs
 
 
 def calculate_cv_recall(
-    results: dict[str, dict[str, float]],
+    results: Mapping[str, Mapping[str, float]],
     qrels: RelevantDocumentsType,
     k_values: list[int],
     skip_first_result: bool = False,

mteb/_evaluators/sklearn_evaluator.py

@@ -1,5 +1,5 @@
 import logging
-from typing import Any, Protocol
+from typing import Any, Protocol, cast
 
 import numpy as np
 from datasets import Dataset
@@ -9,7 +9,7 @@ from typing_extensions import Self
 from mteb._create_dataloaders import create_dataloader
 from mteb.abstasks.task_metadata import TaskMetadata
 from mteb.models import EncoderProtocol
-from mteb.types import BatchedInput
+from mteb.types import Array, BatchedInput
 
 from .evaluator import Evaluator
 
@@ -17,11 +17,11 @@ logger = logging.getLogger(__name__)
 
 
 class SklearnModelProtocol(Protocol):
-    def fit(self, X: np.ndarray, y: np.ndarray | list[int]) -> None: ...  # noqa: N803
-    def predict(self, X: np.ndarray) -> np.ndarray: ...  # noqa: N803
+    def fit(self, X: Array, y: np.ndarray | list[int]) -> None: ...  # noqa: N803
+    def predict(self, X: Array) -> np.ndarray: ...  # noqa: N803
     def get_params(self) -> dict[str, Any]: ...
-    def set_params(self, **kwargs: dict[str, Any]) -> Self: ...
-    def score(self, X: np.ndarray, y: np.ndarray | list[int]) -> float: ...  # noqa: N803
+    def set_params(self, random_state: int, **kwargs: dict[str, Any]) -> Self: ...
+    def score(self, X: Array, y: np.ndarray | list[int]) -> float: ...  # noqa: N803
 
 
 class SklearnEvaluator(Evaluator):
@@ -71,8 +71,8 @@ class SklearnEvaluator(Evaluator):
         model: EncoderProtocol,
         *,
         encode_kwargs: dict[str, Any],
-        test_cache: np.ndarray | None = None,
-    ) -> tuple[np.ndarray, np.ndarray]:
+        test_cache: Array | None = None,
+    ) -> tuple[np.ndarray, Array]:
         """Classification evaluation by training a sklearn classifier on the embeddings of the training set and evaluating on the embeddings of the test set.
 
         Args:
@@ -104,6 +104,7 @@ class SklearnEvaluator(Evaluator):
                 hf_subset=self.hf_subset,
                 **encode_kwargs,
             )
+            test_cache = cast(Array, test_cache)
 
         logger.info("Running - Fitting classifier...")
         y_train = self.train_dataset[self.label_column_name]

mteb/_evaluators/text/bitext_mining_evaluator.py

@@ -1,7 +1,6 @@
 import logging
 from typing import Any
 
-import numpy as np
 import torch
 from datasets import Dataset
 from tqdm.auto import tqdm
@@ -10,6 +9,7 @@ from mteb._create_dataloaders import _create_dataloader_from_texts
 from mteb._evaluators.evaluator import Evaluator
 from mteb.abstasks.task_metadata import TaskMetadata
 from mteb.models import EncoderProtocol
+from mteb.types import Array
 
 logger = logging.getLogger(__name__)
 
@@ -69,11 +69,11 @@ class BitextMiningEvaluator(Evaluator):
 
     def _similarity_search(
         self,
-        query_embeddings: np.ndarray,
-        corpus_embeddings: np.ndarray,
+        query_embeddings: Array,
+        corpus_embeddings: Array,
         model: EncoderProtocol,
         query_chunk_size: int = 100,
-        corpus_chunk_size: int = 500000,
+        corpus_chunk_size: int = 500_000,
     ) -> list[dict[str, float]]:
         """This function performs a cosine similarity search between a list of query embeddings and a list of corpus embeddings.
 
@@ -104,13 +104,15 @@ class BitextMiningEvaluator(Evaluator):
         ):
             query_embeddings = query_embeddings.to(corpus_embeddings.device)
 
-        queries_result_list = [[] for _ in range(len(query_embeddings))]
+        queries_result_list: list[list[dict[str, float]]] = [
+            [] for _ in range(len(query_embeddings))
+        ]
 
         for query_start_idx in range(0, len(query_embeddings), query_chunk_size):
             # Iterate over chunks of the corpus
             for corpus_start_idx in range(0, len(corpus_embeddings), corpus_chunk_size):
                 # Compute cosine similarities
-                similarity_scores = model.similarity(  # type: ignore
+                similarity_scores = model.similarity(
                     query_embeddings[
                         query_start_idx : query_start_idx + query_chunk_size
                     ],
@@ -120,15 +122,17 @@ class BitextMiningEvaluator(Evaluator):
                 )
 
                 # Get top-k scores
-                cos_scores_top_k_values, cos_scores_top_k_idx = torch.topk(
-                    torch.tensor(similarity_scores),
-                    1,
-                    dim=1,
-                    largest=True,
-                    sorted=False,
+                cos_scores_top_k_values_tensor, cos_scores_top_k_idx_tensor = (
+                    torch.topk(
+                        torch.tensor(similarity_scores),
+                        1,
+                        dim=1,
+                        largest=True,
+                        sorted=False,
+                    )
                 )
-                cos_scores_top_k_values = cos_scores_top_k_values.cpu().tolist()
-                cos_scores_top_k_idx = cos_scores_top_k_idx.cpu().tolist()
+                cos_scores_top_k_values = cos_scores_top_k_values_tensor.cpu().tolist()
+                cos_scores_top_k_idx = cos_scores_top_k_idx_tensor.cpu().tolist()
 
                 for query_itr in range(len(similarity_scores)):
                     for sub_corpus_id, score in zip(
@@ -141,11 +145,14 @@ class BitextMiningEvaluator(Evaluator):
                             {"corpus_id": corpus_id, "score": score}
                         )
 
+        result_queries_list: list[dict[str, float]] = [
+            {} for _ in range(len(query_embeddings))
+        ]
         # Sort and strip to top_k results
         for idx in range(len(queries_result_list)):
             queries_result_list[idx] = sorted(
                 queries_result_list[idx], key=lambda x: x["score"], reverse=True
             )
-            queries_result_list[idx] = queries_result_list[idx][0]
+            result_queries_list[idx] = queries_result_list[idx][0]
 
-        return queries_result_list
+        return result_queries_list

mteb/_evaluators/text/summarization_evaluator.py

@@ -135,10 +135,10 @@ class SummarizationEvaluator(Evaluator):
         )
 
         # Split the embeddings into the original human & machine summaries
-        embs_human_summaries_all = np.split(
+        embs_human_summaries_all_split = np.split(
             embs_human_summaries_all, np.cumsum(human_lens)[:-1]
         )
-        embs_machine_summaries_all = np.split(
+        embs_machine_summaries_all_split = np.split(
             embs_machine_summaries_all, np.cumsum(machine_lens)[:-1]
         )
 
@@ -148,7 +148,9 @@ class SummarizationEvaluator(Evaluator):
         all_human_scores = []
 
         for i, (embs_human_summaries, embs_machine_summaries) in tqdm(
-            enumerate(zip(embs_human_summaries_all, embs_machine_summaries_all)),
+            enumerate(
+                zip(embs_human_summaries_all_split, embs_machine_summaries_all_split)
+            ),
             desc="Scoring",
             total=len(self.human_summaries),
         ):
@@ -164,7 +166,7 @@ class SummarizationEvaluator(Evaluator):
                 dot_scores = dot_score(emb_machine_summary, embs_human_summaries)
 
                 _sim_score = [
-                    float(model.similarity(emb_machine_summary, emb_human_summary))  # type: ignore
+                    float(model.similarity(emb_machine_summary, emb_human_summary))
                     for emb_human_summary in embs_human_summaries
                 ]
                 sim_score = torch.tensor(_sim_score)
@@ -216,17 +218,19 @@ class SummarizationEvaluator(Evaluator):
             strict=True,
         ):
             cosine_spearman_scores.append(
-                spearmanr(human_scores, cosine_pred_scores).statistic
+                float(spearmanr(human_scores, cosine_pred_scores).statistic)
             )
             cosine_pearson_scores.append(
-                pearsonr(human_scores, cosine_pred_scores).statistic
+                float(pearsonr(human_scores, cosine_pred_scores).statistic)
             )
             dot_spearman_scores.append(
-                spearmanr(human_scores, dot_pred_scores).statistic
+                float(spearmanr(human_scores, dot_pred_scores).statistic)
+            )
+            dot_pearson_scores.append(
+                float(pearsonr(human_scores, dot_pred_scores).statistic)
             )
-            dot_pearson_scores.append(pearsonr(human_scores, dot_pred_scores).statistic)
-            spearman_scores.append(spearmanr(human_scores, sim_scores).statistic)
-            pearson_scores.append(pearsonr(human_scores, sim_scores).statistic)
+            spearman_scores.append(float(spearmanr(human_scores, sim_scores).statistic))
+            pearson_scores.append(float(pearsonr(human_scores, sim_scores).statistic))
 
         return SummarizationMetrics(
             pearson=float(np.mean(pearson_scores)),
@@ -273,10 +277,10 @@ class DeprecatedSummarizationEvaluator(SummarizationEvaluator):
             pearson_scores.append(pearsonr(human_scores, sim_scores))
 
         return SummarizationMetrics(
-            pearson=float(np.mean(pearson_scores)),
-            spearman=float(np.mean(spearman_scores)),
-            cosine_spearman=float(np.mean(cosine_spearman_scores)),
-            cosine_pearson=float(np.mean(cosine_pearson_scores)),
-            dot_pearson=float(np.mean(dot_pearson_scores)),
-            dot_spearman=float(np.mean(dot_spearman_scores)),
+            pearson=float(np.mean(pearson_scores)),  # type: ignore[arg-type]
+            spearman=float(np.mean(spearman_scores)),  # type: ignore[arg-type]
+            cosine_spearman=float(np.mean(cosine_spearman_scores)),  # type: ignore[arg-type]
+            cosine_pearson=float(np.mean(cosine_pearson_scores)),  # type: ignore[arg-type]
+            dot_pearson=float(np.mean(dot_pearson_scores)),  # type: ignore[arg-type]
+            dot_spearman=float(np.mean(dot_spearman_scores)),  # type: ignore[arg-type]
         )

mteb/abstasks/_data_filter/filters.py

@@ -61,7 +61,7 @@ def filter_unclear_label(
         for text, label in zip(ds[input_column], ds[label_column]):
             key = text.strip().lower()
             normalized.setdefault(key, set()).add(
-                label if isinstance(label, (str, int, float)) else tuple(label)
+                label if isinstance(label, (str, int, float)) else tuple(label)  # type: ignore[arg-type]
             )
 
     bad_texts = {t for t, labels in normalized.items() if len(labels) > 1}

mteb/abstasks/_data_filter/task_pipelines.py

@@ -89,6 +89,9 @@ def process_classification(
             subset=None,
         )
 
+    if task.dataset is None:
+        raise ValueError("Task dataset is None.")
+
     new_ds = {}
     for subset in task.dataset:
         new_ds[subset] = clean_dataset(

mteb/abstasks/_statistics_calculation.py

@@ -2,7 +2,8 @@ from __future__ import annotations
 
 import hashlib
 from collections import Counter
-from typing import TYPE_CHECKING
+from collections.abc import Mapping
+from typing import TYPE_CHECKING, cast
 
 from mteb.types import TopRankedDocumentsType
 from mteb.types.statistics import (
@@ -52,7 +53,7 @@ def calculate_image_statistics(images: list[Image.Image]) -> ImageStatistics:
     seen_hashes: set[str] = set()
 
     for img in images:
-        width, height = img.size  # type: ignore
+        width, height = img.size
         img_heights.append(height)
         img_widths.append(width)
 
@@ -82,17 +83,24 @@ def calculate_label_statistics(labels: list[int | list[int]]) -> LabelStatistics
         LabelStatistics: A dictionary containing the descriptive statistics.
 
     """
+    total_labels: list[int | None] = []
+
     if not isinstance(labels[0], list):
-        label_len = [1] * len(labels)
-        total_label_len = len(labels)
-        total_labels = labels
+        # single label classification
+        single_label = cast(list[int], labels)
+        label_len = [1] * len(single_label)
+        total_label_len = len(single_label)
+        total_labels.extend(single_label)
     elif isinstance(labels[0], list):
         # multilabel classification
-        label_len = [len(l) for l in labels]
+        multilabel_labels = cast(list[list[int]], labels)
+        label_len = [len(l) for l in multilabel_labels]
         total_label_len = sum(label_len)
-        total_labels = []
-        for l in labels:
-            total_labels.extend(l if len(l) > 0 else [None])
+        for l in multilabel_labels:
+            if l and len(l) > 0:
+                total_labels.extend(l)
+            else:
+                total_labels.append(None)
     else:
         raise ValueError(
             "Labels must be a list of integers or a list of lists of integers."
@@ -159,7 +167,7 @@ def calculate_top_ranked_statistics(
 
 
 def calculate_relevant_docs_statistics(
-    relevant_docs: dict[str, dict[str, float]],
+    relevant_docs: Mapping[str, Mapping[str, int]],
 ) -> RelevantDocsStatistics:
     qrels_lengths = [len(relevant_docs[qid]) for qid in relevant_docs]
     unique_qrels = len({doc for qid in relevant_docs for doc in relevant_docs[qid]})

mteb/abstasks/_stratification.py

@@ -39,6 +39,7 @@ Bibtex:
 """
 
 import itertools
+from typing import Any
 
 import numpy as np
 import scipy.sparse as sp
@@ -119,8 +120,10 @@ def _get_most_desired_combination(samples_with_combination: dict):
         if support_size == 0:
             continue
         if currently_chosen is None or (
-            best_number_of_combinations < number_of_combinations  # type: ignore
-            and best_support_size > support_size  # type: ignore
+            best_number_of_combinations is not None
+            and best_support_size is not None
+            and best_number_of_combinations < number_of_combinations
+            and best_support_size > support_size
         ):
             currently_chosen = combination
             best_number_of_combinations, best_support_size = (
@@ -162,7 +165,7 @@ class IterativeStratification(_BaseKFold):
         self._rng_state = check_random_state(random_state)
         need_shuffle = shuffle or random_state is not None
         self.order = order
-        super().__init__(  # type: ignore
+        super().__init__(
             n_splits,
             shuffle=need_shuffle,
             random_state=self._rng_state if need_shuffle else None,
@@ -172,8 +175,7 @@ class IterativeStratification(_BaseKFold):
             self.percentage_per_fold = sample_distribution_per_fold
         else:
             self.percentage_per_fold = [
-                1 / float(self.n_splits)
-                for _ in range(self.n_splits)  # type: ignore
+                1 / float(self.n_splits) for _ in range(self.n_splits)
             ]
 
     def _prepare_stratification(
@@ -182,9 +184,9 @@ class IterativeStratification(_BaseKFold):
         list[list[int]],
         dict[int, bool],
         list[list[int]],
-        list[list[list[int]]],
-        dict[tuple[int, ...], list[int]],
-        list[list[int]],
+        list[list[Any]],
+        dict[str, list[Any]],
+        list[list[Any]],
     ]:
         """Prepares variables for performing stratification
 
@@ -206,14 +208,14 @@ class IterativeStratification(_BaseKFold):
         """
         self.n_samples, self.n_labels = y.shape
         self.desired_samples_per_fold = np.array(
-            [self.percentage_per_fold[i] * self.n_samples for i in range(self.n_splits)]  # type: ignore
+            [self.percentage_per_fold[i] * self.n_samples for i in range(self.n_splits)]
         )
         rows = sp.lil_matrix(y).rows
         rows_used = dict.fromkeys(range(self.n_samples), False)
         all_combinations = []
-        per_row_combinations = [[] for i in range(self.n_samples)]
-        samples_with_combination = {}
-        folds = [[] for _ in range(self.n_splits)]  # type: ignore
+        per_row_combinations: list[list[Any]] = [[] for i in range(self.n_samples)]
+        samples_with_combination: dict[str, list[Any]] = {}
+        folds: list[list[int]] = [[] for _ in range(self.n_splits)]
 
         # for every row
         for sample_index, label_assignment in enumerate(rows):
@@ -229,21 +231,19 @@ class IterativeStratification(_BaseKFold):
                 all_combinations.append(combination)
                 per_row_combinations[sample_index].append(combination)
 
-        all_combinations = [list(x) for x in set(all_combinations)]
-
         self.desired_samples_per_combination_per_fold = {
             combination: np.array(
                 [
                     len(evidence_for_combination) * self.percentage_per_fold[j]
-                    for j in range(self.n_splits)  # type: ignore
+                    for j in range(self.n_splits)
                 ]
             )
             for combination, evidence_for_combination in samples_with_combination.items()
         }
         return (
-            rows,
+            rows.tolist(),
             rows_used,
-            all_combinations,
+            [list(x) for x in set(all_combinations)],
             per_row_combinations,
             samples_with_combination,
             folds,
@@ -328,7 +328,7 @@ class IterativeStratification(_BaseKFold):
             per_row_combinations,
             samples_with_combination,
             folds,
-        ) = self._prepare_stratification(y)  # type: ignore
+        ) = self._prepare_stratification(y)
 
         self._distribute_positive_evidence(
             rows_used, folds, samples_with_combination, per_row_combinations