diff --git a/src/winml/modelkit/datasets/__init__.py b/src/winml/modelkit/datasets/__init__.py
index 66bee45de..b8f9f231a 100644
--- a/src/winml/modelkit/datasets/__init__.py
+++ b/src/winml/modelkit/datasets/__init__.py
@@ -26,6 +26,7 @@
 from .processor_utils import get_image_processor_config
 from .random_dataset import RandomDataset
 from .text import TextDataset
+from .zero_shot_classification import ZeroShotClassificationDataset
 
 
 if TYPE_CHECKING:
@@ -47,7 +48,7 @@
     "sentence-similarity": TextDataset,
     "next-sentence-prediction": TextDataset,
     "fill-mask": TextDataset,
-    "zero-shot-classification": TextDataset,
+    "zero-shot-classification": ZeroShotClassificationDataset,
     "image-segmentation": ImageSegmentationDataset,
     "mask-generation": MaskGenerationDataset,
     "depth-estimation": DepthEstimationDataset,
@@ -337,6 +338,7 @@ def __len__(self) -> int:
     "ObjectDetectionDataset",
     "RandomDataset",
     "TextDataset",
+    "ZeroShotClassificationDataset",
     # Utilities
     "format_data",
     "get_image_processor_config",
diff --git a/src/winml/modelkit/datasets/zero_shot_classification.py b/src/winml/modelkit/datasets/zero_shot_classification.py
new file mode 100644
index 000000000..3991b7f60
--- /dev/null
+++ b/src/winml/modelkit/datasets/zero_shot_classification.py
@@ -0,0 +1,203 @@
+# -------------------------------------------------------------------------
+# Copyright (c) Microsoft Corporation. All rights reserved.
+# Licensed under the MIT License.
+# --------------------------------------------------------------------------
+"""Zero-shot classification calibration dataset.
+
+Zero-shot text classifiers consume natural-language premise/hypothesis pairs.
+This specialization keeps calibration aligned with that input distribution
+instead of reusing generic text-classification defaults.
+"""
+
+from __future__ import annotations
+
+import logging
+from math import ceil
+from numbers import Integral
+from random import Random
+from typing import TYPE_CHECKING, Any, cast
+
+from datasets import load_dataset
+from transformers import AutoTokenizer
+
+from .base import BaseTaskDataset
+
+
+if TYPE_CHECKING:
+    from collections.abc import Sequence
+
+
+logger = logging.getLogger(__name__)
+
+DEFAULT_ZERO_SHOT_CLASSIFICATION_DATASET = "fancyzhx/ag_news"
+DEFAULT_ZERO_SHOT_CLASSIFICATION_SPLIT = "test"
+DEFAULT_ZERO_SHOT_TEXT_COLUMN = "text"
+DEFAULT_ZERO_SHOT_CANDIDATE_LABELS = ("World", "Sports", "Business", "Sci/Tech")
+DEFAULT_ZERO_SHOT_HYPOTHESIS_TEMPLATE = "This text is about {}."
+
+
+class ZeroShotClassificationDataset(BaseTaskDataset):
+    """Calibration dataset for zero-shot text classification models."""
+
+    DEFAULT_SEQ_LEN = 128
+
+    def __init__(
+        self,
+        model_name: str,
+        dataset_name: str | None = None,
+        max_samples: int | None = None,
+        data_split: str | None = None,
+        *,
+        candidate_labels: str | Sequence[str] | None = None,
+        hypothesis_template: str | None = None,
+        input_column: str | None = None,
+        max_length: int | None = None,
+        io_config: dict | None = None,
+        io_mapping: dict | None = None,
+        **kwargs: Any,
+    ) -> None:
+        self._candidate_labels = self._normalize_candidate_labels(candidate_labels)
+        self._hypothesis_template = hypothesis_template or DEFAULT_ZERO_SHOT_HYPOTHESIS_TEMPLATE
+        self._input_column = input_column or DEFAULT_ZERO_SHOT_TEXT_COLUMN
+        self._max_length = max_length
+        self._io_config = io_config
+        self._io_mapping = io_mapping or {}
+
+        super().__init__(
+            model_name=model_name,
+            dataset_name=dataset_name,
+            max_samples=max_samples,
+            data_split=data_split,
+            io_config=io_config,
+            io_mapping=io_mapping,
+            **kwargs,
+        )
+
+    @staticmethod
+    def _normalize_candidate_labels(candidate_labels: str | Sequence[str] | None) -> list[str]:
+        """Return non-empty labels from comma-separated or sequence input."""
+        if candidate_labels is None:
+            raw_labels: Sequence[str] = DEFAULT_ZERO_SHOT_CANDIDATE_LABELS
+        elif isinstance(candidate_labels, str):
+            raw_labels = candidate_labels.split(",")
+        else:
+            raw_labels = candidate_labels
+
+        labels = [str(label).strip() for label in raw_labels if str(label).strip()]
+        if not labels:
+            raise ValueError("candidate_labels must contain at least one label")
+        return labels
+
+    def _get_default_dataset(self) -> None:
+        """Set the built-in zero-shot dataset defaults."""
+        if self._dataset_name is None:
+            self._dataset_name = DEFAULT_ZERO_SHOT_CLASSIFICATION_DATASET
+        if self._data_split is None:
+            self._data_split = DEFAULT_ZERO_SHOT_CLASSIFICATION_SPLIT
+
+    def _resolve_max_length(self) -> None:
+        """Resolve tokenizer max length from ONNX io_config when available."""
+        if self._max_length is None:
+            self._max_length = self.DEFAULT_SEQ_LEN
+
+        if not self._io_config:
+            return
+
+        onnx_name = self._io_mapping.get("input_ids", "input_ids")
+        input_config = self._io_config.get(onnx_name)
+        if not input_config:
+            return
+
+        shape = input_config.get("shape", [])
+        if len(shape) > 1 and isinstance(shape[1], Integral):
+            self._max_length = int(shape[1])
+            logger.info("max_length=%d from io_config[%s]", self._max_length, onnx_name)
+
+    def _source_sample_count(self) -> int | None:
+        """Number of source rows needed to produce max_samples calibration pairs."""
+        if self._max_samples is None:
+            return None
+        return ceil(self._max_samples / len(self._candidate_labels))
+
+    def _load_and_sample(self) -> Any:
+        """Load the source text dataset and cap rows before pair expansion."""
+        subset = self._config.get("subset") or self._config.get("dataset_config_name")
+        load_args = [self._dataset_name]
+        if subset:
+            load_args.append(subset)
+
+        logger.info(
+            "Loading zero-shot calibration dataset: %s (split=%s)",
+            load_args,
+            self._data_split,
+        )
+        dataset = load_dataset(*load_args, split=self._data_split)
+
+        source_count = self._source_sample_count()
+        shuffle = self._config.get("shuffle", False)
+        seed = self._config.get("seed", 42)
+
+        if source_count is not None:
+            n = min(source_count, len(dataset))
+            indices = Random(seed).sample(range(len(dataset)), n) if shuffle else list(range(n))
+            dataset = dataset.select(indices)
+        elif shuffle:
+            dataset = dataset.shuffle(seed=seed)
+
+        return dataset
+
+    def _apply_io_mapping(self, sample: dict[str, Any]) -> dict[str, Any]:
+        """Rename tokenizer output fields to ONNX input names when requested."""
+        if not self._io_mapping:
+            return sample
+        return {self._io_mapping.get(key, key): value for key, value in sample.items()}
+
+    def _initialize(self) -> None:
+        """Load text rows and tokenize them as zero-shot premise/hypothesis pairs."""
+        self._get_default_dataset()
+        self._resolve_max_length()
+        dataset = self._load_and_sample()
+        tokenizer = AutoTokenizer.from_pretrained(self._model_name, use_fast=True)
+
+        samples: list[dict[str, Any]] = []
+        for example in dataset:
+            if self._input_column not in example:
+                raise ValueError(
+                    f"Column '{self._input_column}' not found in {self._dataset_name}; "
+                    f"available columns: {sorted(example)}"
+                )
+            premise = str(example[self._input_column])
+            for label in self._candidate_labels:
+                hypothesis = self._hypothesis_template.format(label)
+                tokenized = dict(
+                    tokenizer(
+                        premise,
+                        hypothesis,
+                        padding="max_length",
+                        truncation=True,
+                        max_length=self._max_length,
+                        return_tensors="pt",
+                    )
+                )
+                samples.append(self._apply_io_mapping(tokenized))
+                if self._max_samples is not None and len(samples) >= self._max_samples:
+                    self._dataset = samples
+                    return
+
+        self._dataset = samples
+        logger.info("Initialized zero-shot calibration dataset with %d pairs", len(samples))
+
+    def __getitem__(self, idx: int) -> dict[str, Any]:
+        """Get a tokenized premise/hypothesis calibration sample."""
+        return cast("dict[str, Any]", self._dataset[idx])
+
+    @property
+    def label_col(self) -> str:
+        """Zero-shot calibration does not consume dataset labels."""
+        return ""
+
+    @property
+    def max_length(self) -> int:
+        """Sequence length."""
+        assert self._max_length is not None, "max_length not resolved"
+        return self._max_length
diff --git a/tests/integration/datasets/test_zero_shot_classification.py b/tests/integration/datasets/test_zero_shot_classification.py
new file mode 100644
index 000000000..b82ad25ea
--- /dev/null
+++ b/tests/integration/datasets/test_zero_shot_classification.py
@@ -0,0 +1,89 @@
+# -------------------------------------------------------------------------
+# Copyright (c) Microsoft Corporation. All rights reserved.
+# Licensed under the MIT License.
+# --------------------------------------------------------------------------
+"""Tests for zero-shot-classification calibration datasets."""
+
+from __future__ import annotations
+
+from typing import Any
+from unittest.mock import MagicMock, patch
+
+import torch
+from datasets.features import ClassLabel, Value
+
+
+class _FakeDataset:
+    def __init__(self, rows: list[dict[str, Any]]) -> None:
+        self._rows = rows
+        self.features = {
+            "text": Value("string"),
+            "label": ClassLabel(names=["World", "Sports", "Business", "Sci/Tech"]),
+        }
+
+    def __len__(self) -> int:
+        return len(self._rows)
+
+    def __getitem__(self, idx: int) -> dict[str, Any]:
+        return self._rows[idx]
+
+    def select(self, indices: list[int]) -> _FakeDataset:
+        return _FakeDataset([self._rows[i] for i in indices])
+
+
+class _FakeTokenizer:
+    def __init__(self) -> None:
+        self.calls: list[tuple[str, str]] = []
+
+    def __call__(self, premise: str, hypothesis: str, **_: Any) -> dict[str, torch.Tensor]:
+        self.calls.append((premise, hypothesis))
+        return {
+            "input_ids": torch.ones(1, 8, dtype=torch.int64),
+            "attention_mask": torch.ones(1, 8, dtype=torch.int64),
+        }
+
+
+class TestZeroShotClassificationDataset:
+    @patch("winml.modelkit.datasets.zero_shot_classification.load_dataset")
+    @patch("winml.modelkit.datasets.zero_shot_classification.AutoTokenizer")
+    def test_defaults_use_ag_news_sentence_pairs(
+        self,
+        mock_tokenizer_cls: MagicMock,
+        mock_load_dataset: MagicMock,
+    ) -> None:
+        from winml.modelkit.datasets import ZeroShotClassificationDataset
+        from winml.modelkit.datasets.zero_shot_classification import (
+            DEFAULT_ZERO_SHOT_CLASSIFICATION_DATASET,
+            DEFAULT_ZERO_SHOT_CLASSIFICATION_SPLIT,
+        )
+
+        tokenizer = _FakeTokenizer()
+        mock_tokenizer_cls.from_pretrained.return_value = tokenizer
+        mock_load_dataset.return_value = _FakeDataset(
+            [{"text": "Markets rallied after the earnings report.", "label": 2}]
+        )
+
+        dataset = ZeroShotClassificationDataset(
+            model_name="cross-encoder/nli-deberta-v3-small",
+            max_samples=4,
+        )
+
+        assert dataset.dataset_name == DEFAULT_ZERO_SHOT_CLASSIFICATION_DATASET
+        assert dataset.data_split == DEFAULT_ZERO_SHOT_CLASSIFICATION_SPLIT
+        mock_load_dataset.assert_called_once_with(
+            DEFAULT_ZERO_SHOT_CLASSIFICATION_DATASET,
+            split=DEFAULT_ZERO_SHOT_CLASSIFICATION_SPLIT,
+        )
+        assert tokenizer.calls == [
+            ("Markets rallied after the earnings report.", "This text is about World."),
+            ("Markets rallied after the earnings report.", "This text is about Sports."),
+            ("Markets rallied after the earnings report.", "This text is about Business."),
+            ("Markets rallied after the earnings report.", "This text is about Sci/Tech."),
+        ]
+        assert len(dataset) == 4
+        assert set(dataset[0]) == {"input_ids", "attention_mask"}
+
+    def test_zero_shot_task_uses_specialized_dataset(self) -> None:
+        from winml.modelkit.datasets import TASK_DATASET_MAPPING, ZeroShotClassificationDataset
+
+        assert TASK_DATASET_MAPPING["zero-shot-classification"] is ZeroShotClassificationDataset