fix(datasets): use fully qualified HF dataset ids

examples/recipes/cardiffnlp_twitter-roberta-base-sentiment-latest/text-classification_fp16_config.json

@@ -53,7 +53,7 @@
   "eval": {
     "task": "text-classification",
     "dataset": {
-      "path": "tweet_eval",
+      "path": "cardiffnlp/tweet_eval",
       "name": "sentiment",
       "samples": 100,
       "columns_mapping": {

examples/recipes/cardiffnlp_twitter-roberta-base-sentiment-latest/text-classification_w8a16_config.json

@@ -70,7 +70,7 @@
   "eval": {
     "task": "text-classification",
     "dataset": {
-      "path": "tweet_eval",
+      "path": "cardiffnlp/tweet_eval",
       "name": "sentiment",
       "samples": 1000,
       "columns_mapping": {

scripts/e2e_eval/cache/baseline_cache.json

@@ -9,15 +9,15 @@
     "elapsed": 59.2,
     "command": "python.exe run_pytorch_baseline.py --model finbert --task text-classification --device cpu --num-samples 1000 --dataset finbert_dataset --split val"
   },
-  "cardiffnlp/twitter-roberta-base-sentiment-latest|text-classification|tweet_eval|sentiment||1000": {
+  "cardiffnlp/twitter-roberta-base-sentiment-latest|text-classification|cardiffnlp/tweet_eval|sentiment||1000": {
     "status": "PASS",
     "metric": {
       "metric": "accuracy",
       "value": 0.769,
       "num_samples": 1000
     },
     "elapsed": 91.5,
-    "command": "python.exe run_pytorch_baseline.py --model twitter-roberta-base-sentiment-latest --task text-classification --device cpu --num-samples 1000 --dataset tweet_eval --dataset-config sentiment --columns-mapping {\"input_column\": \"text\"}"
+    "command": "python.exe run_pytorch_baseline.py --model twitter-roberta-base-sentiment-latest --task text-classification --device cpu --num-samples 1000 --dataset cardiffnlp/tweet_eval --dataset-config sentiment --columns-mapping {\"input_column\": \"text\"}"
   },
   "distilbert/distilbert-base-uncased-finetuned-sst-2-english|text-classification|nyu-mll/glue|sst2||1000": {
     "status": "PASS",
@@ -959,15 +959,15 @@
     "elapsed": 52.0,
     "command": "python.exe run_pytorch_baseline.py --model finbert --task text-classification --device cpu --num-samples 100 --dataset finbert_dataset --split val --winml-metric-key accuracy"
   },
-  "cardiffnlp/twitter-roberta-base-sentiment-latest|text-classification|tweet_eval|sentiment||100": {
+  "cardiffnlp/twitter-roberta-base-sentiment-latest|text-classification|cardiffnlp/tweet_eval|sentiment||100": {
     "status": "PASS",
     "metric": {
       "metric": "accuracy",
       "value": 0.81,
       "num_samples": 100
     },
     "elapsed": 54.9,
-    "command": "python.exe run_pytorch_baseline.py --model twitter-roberta-base-sentiment-latest --task text-classification --device cpu --num-samples 100 --dataset tweet_eval --dataset-config sentiment --columns-mapping {\"input_column\": \"text\"} --winml-metric-key accuracy"
+    "command": "python.exe run_pytorch_baseline.py --model twitter-roberta-base-sentiment-latest --task text-classification --device cpu --num-samples 100 --dataset cardiffnlp/tweet_eval --dataset-config sentiment --columns-mapping {\"input_column\": \"text\"} --winml-metric-key accuracy"
   },
   "distilbert/distilbert-base-uncased-finetuned-sst-2-english|text-classification|nyu-mll/glue|sst2||100": {
     "status": "PASS",

scripts/e2e_eval/testsets/models_with_acc.json

@@ -18,7 +18,7 @@
     "group": "Top200",
     "priority": "P1",
     "dataset_config": {
-      "path": "tweet_eval",
+      "path": "cardiffnlp/tweet_eval",
       "name": "sentiment",
       "metric": "accuracy",
       "columns_mapping": {

src/winml/modelkit/commands/eval.py

@@ -45,7 +45,7 @@
     "dataset_path",
     type=str,
     default=None,
-    help="HF dataset path (e.g. 'imagenet-1k', 'glue'). "
+    help="HF dataset path (e.g. 'imagenet-1k', 'nyu-mll/glue'). "
     "If omitted, uses a default dataset for the task.",
 )
 @click.option(

src/winml/modelkit/datasets/text.py

@@ -27,6 +27,11 @@
 
 logger = logging.getLogger(__name__)
 
+# HF requires fully qualified repository ids ("namespace/name"); the legacy
+# canonical alias "glue" is no longer resolvable.
+DEFAULT_TEXT_DATASET = "nyu-mll/glue"
+DEFAULT_TEXT_DATASET_SUBSET = "mrpc"
+
 
 class TextDataset(BaseTaskDataset):
     """Dataset for text tasks with universal tokenization.
@@ -56,7 +61,7 @@ def __init__(
 
         Args:
             model_name: HuggingFace model identifier
-            dataset_name: Dataset name (default: glue)
+            dataset_name: Dataset name (default: nyu-mll/glue)
             max_samples: Maximum samples (None = use all)
             data_split: Dataset split (default: train)
             max_length: Sequence length (default: from io_config or 128)
@@ -85,8 +90,8 @@ def __init__(
     def _get_default_dataset(self) -> None:
         """Set default dataset if none specified."""
         if self._dataset_name is None:
-            self._dataset_name = "glue"
-            self._config["subset"] = self._config.get("subset", "mrpc")
+            self._dataset_name = DEFAULT_TEXT_DATASET
+            self._config["subset"] = self._config.get("subset", DEFAULT_TEXT_DATASET_SUBSET)
             self._data_split = self._data_split or "train"
 
     def _resolve_max_length(self) -> None:

src/winml/modelkit/eval/config.py

@@ -20,7 +20,7 @@ class DatasetConfig:
     """Dataset configuration, aligned with HF load_dataset() API.
 
     Attributes:
-        path: HF dataset path (e.g., "imagenet-1k", "glue").
+        path: HF dataset path (e.g., "imagenet-1k", "nyu-mll/glue").
         name: Config name for multi-config datasets (e.g., "mrpc").
         split: Dataset split.
         samples: Number of samples to evaluate.

tests/integration/datasets/test_text_classification.py

@@ -44,15 +44,15 @@ def test_default_seq_len_constant(self):
         assert TextDataset.DEFAULT_SEQ_LEN == 128
 
     def test_default_dataset_glue_mrpc(self):
-        """Test default dataset is glue/mrpc when none specified."""
+        """Test default dataset is nyu-mll/glue with the mrpc subset."""
         from winml.modelkit.datasets import TextDataset
 
         dataset = TextDataset(
             model_name="bert-base-uncased",
             max_samples=5,
         )
 
-        assert dataset.dataset_name == "glue"
+        assert dataset.dataset_name == "nyu-mll/glue"
         assert dataset.data_split == "train"
 
     def test_explicit_dataset_name(self):
@@ -61,13 +61,13 @@ def test_explicit_dataset_name(self):
 
         dataset = TextDataset(
             model_name="bert-base-uncased",
-            dataset_name="glue",
+            dataset_name="nyu-mll/glue",
             data_split="validation",
             max_samples=5,
             subset="sst2",
         )
 
-        assert dataset.dataset_name == "glue"
+        assert dataset.dataset_name == "nyu-mll/glue"
         assert dataset.data_split == "validation"
 
     def test_max_samples_limits_dataset_size(self):
@@ -274,7 +274,7 @@ def test_single_sentence_detection_sst2(self):
         # GLUE/SST2 is a single sentence task
         dataset = TextDataset(
             model_name="bert-base-uncased",
-            dataset_name="glue",
+            dataset_name="nyu-mll/glue",
             data_split="train",
             max_samples=5,
             subset="sst2",

tests/unit/eval/test_eval.py

@@ -752,7 +752,7 @@ def _fake_compute(**kw):
             model_id="test/model",
             task="text-classification",
             dataset=DatasetConfig(
-                path="glue",
+                path="nyu-mll/glue",
                 name="mrpc",
                 columns_mapping={"input_column": "sentence1", "second_input_column": "sentence2"},
             ),
@@ -803,7 +803,7 @@ def test_sets_padding_for_text_model(
         config = WinMLEvaluationConfig(
             model_id="test/model",
             task="text-classification",
-            dataset=DatasetConfig(path="glue", name="mrpc"),
+            dataset=DatasetConfig(path="nyu-mll/glue", name="mrpc"),
         )
 
         WinMLTextClassificationEvaluator(config, model).compute()
@@ -847,7 +847,7 @@ def test_no_padding_without_tokenizer(
         config = WinMLEvaluationConfig(
             model_id="test/model",
             task="text-classification",
-            dataset=DatasetConfig(path="glue"),
+            dataset=DatasetConfig(path="nyu-mll/glue"),
         )
 
         WinMLTextClassificationEvaluator(config, model).compute()