From 01c440be060ee3315fc98669f164f73eaef6a238 Mon Sep 17 00:00:00 2001
From: ccyyy1023 <2637034749@qq.com>
Date: Thu, 30 Jul 2026 10:02:34 +0800
Subject: [PATCH 1/3] Support GlotLID language and script labels

Signed-off-by: ccyyy1023 <2637034749@qq.com>
---
 .../language-management/language.mdx          | 30 ++++++---
 .../text/filters/fasttext/fasttext_filters.py | 24 ++++++--
 tests/stages/text/filters/__init__.py         | 13 ++++
 .../stages/text/filters/fasttext/__init__.py  | 13 ++++
 .../filters/fasttext/test_fasttext_filters.py | 61 +++++++++++++++++++
 5 files changed, 128 insertions(+), 13 deletions(-)
 create mode 100644 tests/stages/text/filters/__init__.py
 create mode 100644 tests/stages/text/filters/fasttext/__init__.py
 create mode 100644 tests/stages/text/filters/fasttext/test_fasttext_filters.py

diff --git a/fern/versions/main/pages/curate-text/process-data/language-management/language.mdx b/fern/versions/main/pages/curate-text/process-data/language-management/language.mdx
index 8a4714870a..ed5be56ef1 100644
--- a/fern/versions/main/pages/curate-text/process-data/language-management/language.mdx
+++ b/fern/versions/main/pages/curate-text/process-data/language-management/language.mdx
@@ -1,7 +1,7 @@
 ---
-description: "Identify document languages accurately using FastText models supporting 176 languages for multilingual text processing"
+description: "Identify document languages and scripts using FastText and GlotLID models for multilingual text processing"
 categories: ["how-to-guides"]
-tags: ["language-identification", "fasttext", "multilingual", "176-languages", "detection", "classification"]
+tags: ["language-identification", "fasttext", "glotlid", "multilingual", "detection", "classification"]
 personas: ["data-scientist-focused", "mle-focused"]
 difficulty: "intermediate"
 content_type: "how-to"
@@ -18,9 +18,9 @@ NeMo Curator's language identification system works through a three-step process
 
 1. **Text Preprocessing**: For FastText classification, normalize input text by stripping whitespace and converting newlines to spaces.
 
-2. **FastText Language Detection**: The pre-trained FastText language identification model ([`lid.176.bin`](https://fasttext.cc/docs/en/language-identification.html)) analyzes the preprocessed text and returns:
+2. **FastText Language Detection**: A FastText-compatible language identification model, such as [`lid.176.bin`](https://fasttext.cc/docs/en/language-identification.html) or [GlotLID](https://huggingface.co/cis-lmu/glotlid), analyzes the preprocessed text and returns:
    - A confidence score (0.0 to 1.0) indicating certainty of the prediction
-   - A language code (for example, "EN", "ES", "FR") in FastText's two-letter uppercase format
+   - A language code (for example, `en`) or a GlotLID language-script code (for example, `eng_Latn`)
 
 3. **Filtering and Scoring**: The pipeline filters documents based on a configurable confidence threshold (`min_langid_score`) and stores both the confidence score and language code as metadata.
 
@@ -30,11 +30,11 @@ The `FastTextLangId` filter implements this workflow by:
 
 - Loading the FastText language identification model on worker initialization
 - Processing text through `model.predict()` with `k=1` to get the top language prediction
-- Extracting the language code from FastText labels (for example, `__label__en` becomes "EN")
+- Removing the `__label__` prefix while preserving the complete model label (for example, `__label__en` becomes `en` and `__label__eng_Latn` becomes `eng_Latn`)
 - Comparing confidence scores against the threshold to determine document retention
 - Returning results as `[confidence_score, language_code]` for downstream processing
 
-This approach supports **176 languages** with high accuracy, making it suitable for large-scale multilingual dataset curation where language-specific processing and monolingual dataset creation are critical.
+The standard FastText model supports **176 languages**. GlotLID can be used when broader language coverage or script identification is required.
 
 ## Usage
 
@@ -110,13 +110,29 @@ if __name__ == "__main__":
 </Tab>
 </Tabs>
 
+### Using GlotLID
+
+Download the GlotLID FastText model from [Hugging Face](https://huggingface.co/cis-lmu/glotlid), then pass its local path to `FastTextLangId` in the same way as the standard FastText model:
+
+```python
+glotlid_model_path = "/path/to/glotlid/model.bin"
+
+# Keep every script predicted for English.
+english_filter = FastTextLangId(model_path=glotlid_model_path, lang="eng")
+
+# Keep only English written in the Latin script.
+english_latin_filter = FastTextLangId(model_path=glotlid_model_path, lang="eng_Latn")
+```
+
+Language matching is case-insensitive. A filter without an underscore matches the language portion of a GlotLID label, while a filter containing an underscore matches the complete language-script label.
+
 ## Understanding Results
 
 The language identification process adds a score field to each document batch:
 
 1. **`language` field**: Contains the FastText language identification results as a string representation of a list with two elements (for backend compatibility):
    - Element 0: The confidence score (between 0 and 1)
-   - Element 1: The language code in FastText format (for example, "EN" for English, "ES" for Spanish)
+   - Element 1: The complete model label without the `__label__` prefix (for example, `en` or `eng_Latn`)
 
 2. **Task-based processing**: Curator processes documents in batches (tasks), and results are available through the task's Pandas DataFrame:
 
diff --git a/nemo_curator/stages/text/filters/fasttext/fasttext_filters.py b/nemo_curator/stages/text/filters/fasttext/fasttext_filters.py
index 81e8987621..6e6619fdef 100644
--- a/nemo_curator/stages/text/filters/fasttext/fasttext_filters.py
+++ b/nemo_curator/stages/text/filters/fasttext/fasttext_filters.py
@@ -56,12 +56,20 @@ def keep_document(self, score: float) -> bool:
 
 
 class FastTextLangId(DocumentFilter):
+    """Identify and optionally filter languages predicted by a FastText model.
+
+    Language labels may be simple codes such as ``en`` or language-script
+    combinations such as ``eng_Latn``. A language-only filter matches every
+    script for that language, while a language-script filter requires an exact
+    match. Matching is case-insensitive.
+    """
+
     def __init__(self, model_path: str | None = None, min_langid_score: float = 0.3, lang: str | None = None):
         if model_path is None:
             msg = "Must provide a valid path to a FastText model to identify languages with this filter"
             raise ValueError(msg)
         self._model_path = model_path
-        self._lang_code = lang.upper() if lang else None
+        self._lang_code = lang.casefold() if lang else None
         self._cutoff = min_langid_score
         self._name = "lang_id"
 
@@ -73,14 +81,14 @@ def model_check_or_download(self) -> None:
     def load_model(self) -> None:
         self._fasttext_langid_model = fasttext.load_model(self._model_path)
 
-    def score_document(self, text: str) -> list[float | str]:
+    def score_document(self, text: str) -> str:
         # See setup() function in modules/filter.py
         model = self._fasttext_langid_model
 
         pp = text.strip().replace("\n", " ")
         label, score = model.predict([pp], k=1)
         score = score[0][0].item()
-        lang_code = label[0][0][-2:].upper()
+        lang_code = label[0][0].removeprefix("__label__")
 
         # Need to convert it to a string to allow backend conversions
         return str([score, lang_code])
@@ -89,10 +97,14 @@ def keep_document(self, score: float | str) -> bool:
         if isinstance(score, str):
             score_lang = eval(score)  # noqa: S307
             score = score_lang[0]
-            lang = score_lang[1].upper()
-        else :
+            lang = score_lang[1].casefold()
+        else:
             msg = "score must be a string convertible to list"
             raise TypeError(msg)
         if self._lang_code:
-            return score >= self._cutoff and lang == self._lang_code
+            if "_" in self._lang_code:
+                language_matches = lang == self._lang_code
+            else:
+                language_matches = lang.split("_", maxsplit=1)[0] == self._lang_code
+            return score >= self._cutoff and language_matches
         return score >= self._cutoff
diff --git a/tests/stages/text/filters/__init__.py b/tests/stages/text/filters/__init__.py
new file mode 100644
index 0000000000..4fc25d0d3c
--- /dev/null
+++ b/tests/stages/text/filters/__init__.py
@@ -0,0 +1,13 @@
+# Copyright (c) 2026, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
diff --git a/tests/stages/text/filters/fasttext/__init__.py b/tests/stages/text/filters/fasttext/__init__.py
new file mode 100644
index 0000000000..4fc25d0d3c
--- /dev/null
+++ b/tests/stages/text/filters/fasttext/__init__.py
@@ -0,0 +1,13 @@
+# Copyright (c) 2026, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
diff --git a/tests/stages/text/filters/fasttext/test_fasttext_filters.py b/tests/stages/text/filters/fasttext/test_fasttext_filters.py
new file mode 100644
index 0000000000..ae39d0e06e
--- /dev/null
+++ b/tests/stages/text/filters/fasttext/test_fasttext_filters.py
@@ -0,0 +1,61 @@
+# Copyright (c) 2026, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from unittest.mock import Mock
+
+import numpy as np
+import pytest
+
+from nemo_curator.stages.text.filters.fasttext import FastTextLangId
+
+
+@pytest.mark.parametrize(
+    ("label", "expected_language"),
+    [
+        ("__label__en", "en"),
+        ("__label__eng_Latn", "eng_Latn"),
+    ],
+)
+def test_score_document_preserves_complete_fasttext_label(label: str, expected_language: str) -> None:
+    lang_id = FastTextLangId(model_path="model.bin")
+    lang_id._fasttext_langid_model = Mock()
+    lang_id._fasttext_langid_model.predict.return_value = ([[label]], [np.array([0.9])])
+
+    assert lang_id.score_document("Hello, world!") == str([0.9, expected_language])
+
+
+@pytest.mark.parametrize(
+    ("language_filter", "prediction", "expected"),
+    [
+        ("en", "en", True),
+        ("EN", "en", True),
+        ("eng", "eng_Latn", True),
+        ("eng_Latn", "eng_Latn", True),
+        ("ENG_LATN", "eng_Latn", True),
+        ("eng_Cyrl", "eng_Latn", False),
+        ("deu", "eng_Latn", False),
+    ],
+)
+def test_keep_document_filters_language_or_language_script(
+    language_filter: str, prediction: str, expected: bool
+) -> None:
+    lang_id = FastTextLangId(model_path="model.bin", lang=language_filter)
+
+    assert lang_id.keep_document(str([0.9, prediction])) is expected
+
+
+def test_keep_document_applies_score_cutoff_with_glotlid_label() -> None:
+    lang_id = FastTextLangId(model_path="model.bin", min_langid_score=0.8, lang="eng")
+
+    assert not lang_id.keep_document(str([0.7, "eng_Latn"]))

From d545f4502365ab7f69170e22d9bf2d4a9fc726d7 Mon Sep 17 00:00:00 2001
From: ccyyy1023 <2637034749@qq.com>
Date: Thu, 30 Jul 2026 10:23:27 +0800
Subject: [PATCH 2/3] Preserve standard FastText label casing

Signed-off-by: ccyyy1023 <2637034749@qq.com>
---
 .../process-data/language-management/language.mdx           | 6 +++---
 .../stages/text/filters/fasttext/fasttext_filters.py        | 4 +++-
 tests/stages/text/filters/fasttext/test_fasttext_filters.py | 3 ++-
 3 files changed, 8 insertions(+), 5 deletions(-)

diff --git a/fern/versions/main/pages/curate-text/process-data/language-management/language.mdx b/fern/versions/main/pages/curate-text/process-data/language-management/language.mdx
index ed5be56ef1..918d2b8f1f 100644
--- a/fern/versions/main/pages/curate-text/process-data/language-management/language.mdx
+++ b/fern/versions/main/pages/curate-text/process-data/language-management/language.mdx
@@ -20,7 +20,7 @@ NeMo Curator's language identification system works through a three-step process
 
 2. **FastText Language Detection**: A FastText-compatible language identification model, such as [`lid.176.bin`](https://fasttext.cc/docs/en/language-identification.html) or [GlotLID](https://huggingface.co/cis-lmu/glotlid), analyzes the preprocessed text and returns:
    - A confidence score (0.0 to 1.0) indicating certainty of the prediction
-   - A language code (for example, `en`) or a GlotLID language-script code (for example, `eng_Latn`)
+   - An uppercase language code (for example, `EN`) or a GlotLID language-script code (for example, `eng_Latn`)
 
 3. **Filtering and Scoring**: The pipeline filters documents based on a configurable confidence threshold (`min_langid_score`) and stores both the confidence score and language code as metadata.
 
@@ -30,7 +30,7 @@ The `FastTextLangId` filter implements this workflow by:
 
 - Loading the FastText language identification model on worker initialization
 - Processing text through `model.predict()` with `k=1` to get the top language prediction
-- Removing the `__label__` prefix while preserving the complete model label (for example, `__label__en` becomes `en` and `__label__eng_Latn` becomes `eng_Latn`)
+- Removing the `__label__` prefix while preserving standard FastText's uppercase output and complete GlotLID labels (for example, `__label__en` becomes `EN` and `__label__eng_Latn` becomes `eng_Latn`)
 - Comparing confidence scores against the threshold to determine document retention
 - Returning results as `[confidence_score, language_code]` for downstream processing
 
@@ -132,7 +132,7 @@ The language identification process adds a score field to each document batch:
 
 1. **`language` field**: Contains the FastText language identification results as a string representation of a list with two elements (for backend compatibility):
    - Element 0: The confidence score (between 0 and 1)
-   - Element 1: The complete model label without the `__label__` prefix (for example, `en` or `eng_Latn`)
+   - Element 1: The standard uppercase language code or complete GlotLID label without the `__label__` prefix (for example, `EN` or `eng_Latn`)
 
 2. **Task-based processing**: Curator processes documents in batches (tasks), and results are available through the task's Pandas DataFrame:
 
diff --git a/nemo_curator/stages/text/filters/fasttext/fasttext_filters.py b/nemo_curator/stages/text/filters/fasttext/fasttext_filters.py
index 6e6619fdef..9e678adb6e 100644
--- a/nemo_curator/stages/text/filters/fasttext/fasttext_filters.py
+++ b/nemo_curator/stages/text/filters/fasttext/fasttext_filters.py
@@ -58,7 +58,7 @@ def keep_document(self, score: float) -> bool:
 class FastTextLangId(DocumentFilter):
     """Identify and optionally filter languages predicted by a FastText model.
 
-    Language labels may be simple codes such as ``en`` or language-script
+    Language labels may be simple codes such as ``EN`` or language-script
     combinations such as ``eng_Latn``. A language-only filter matches every
     script for that language, while a language-script filter requires an exact
     match. Matching is case-insensitive.
@@ -89,6 +89,8 @@ def score_document(self, text: str) -> str:
         label, score = model.predict([pp], k=1)
         score = score[0][0].item()
         lang_code = label[0][0].removeprefix("__label__")
+        if "_" not in lang_code:
+            lang_code = lang_code.upper()
 
         # Need to convert it to a string to allow backend conversions
         return str([score, lang_code])
diff --git a/tests/stages/text/filters/fasttext/test_fasttext_filters.py b/tests/stages/text/filters/fasttext/test_fasttext_filters.py
index ae39d0e06e..48f3942b13 100644
--- a/tests/stages/text/filters/fasttext/test_fasttext_filters.py
+++ b/tests/stages/text/filters/fasttext/test_fasttext_filters.py
@@ -23,7 +23,7 @@
 @pytest.mark.parametrize(
     ("label", "expected_language"),
     [
-        ("__label__en", "en"),
+        ("__label__en", "EN"),
         ("__label__eng_Latn", "eng_Latn"),
     ],
 )
@@ -39,6 +39,7 @@ def test_score_document_preserves_complete_fasttext_label(label: str, expected_l
     ("language_filter", "prediction", "expected"),
     [
         ("en", "en", True),
+        ("en", "EN", True),
         ("EN", "en", True),
         ("eng", "eng_Latn", True),
         ("eng_Latn", "eng_Latn", True),

From fa849dbd5074f2d6851082c6ab81eed22f8f91d3 Mon Sep 17 00:00:00 2001
From: ccyyy1023 <2637034749@qq.com>
Date: Tue, 4 Aug 2026 10:52:08 +0800
Subject: [PATCH 3/3] Address FastText language label review feedback

Signed-off-by: ccyyy1023 <2637034749@qq.com>
---
 .../language-management/language.mdx          |  6 +-
 .../text/filters/fasttext/fasttext_filters.py | 11 ++-
 tests/stages/text/filters/__init__.py         | 13 ---
 .../stages/text/filters/fasttext/__init__.py  | 13 ---
 .../filters/fasttext/test_fasttext_filters.py | 89 +++++++++++++++++-
 tests/stages/text/modules/test_filters.py     | 90 ++-----------------
 6 files changed, 102 insertions(+), 120 deletions(-)

diff --git a/fern/versions/main/pages/curate-text/process-data/language-management/language.mdx b/fern/versions/main/pages/curate-text/process-data/language-management/language.mdx
index 918d2b8f1f..aa5e911afc 100644
--- a/fern/versions/main/pages/curate-text/process-data/language-management/language.mdx
+++ b/fern/versions/main/pages/curate-text/process-data/language-management/language.mdx
@@ -20,7 +20,7 @@ NeMo Curator's language identification system works through a three-step process
 
 2. **FastText Language Detection**: A FastText-compatible language identification model, such as [`lid.176.bin`](https://fasttext.cc/docs/en/language-identification.html) or [GlotLID](https://huggingface.co/cis-lmu/glotlid), analyzes the preprocessed text and returns:
    - A confidence score (0.0 to 1.0) indicating certainty of the prediction
-   - An uppercase language code (for example, `EN`) or a GlotLID language-script code (for example, `eng_Latn`)
+   - The model's language label (for example, `en`) or language-script label (for example, `eng_Latn`)
 
 3. **Filtering and Scoring**: The pipeline filters documents based on a configurable confidence threshold (`min_langid_score`) and stores both the confidence score and language code as metadata.
 
@@ -30,7 +30,7 @@ The `FastTextLangId` filter implements this workflow by:
 
 - Loading the FastText language identification model on worker initialization
 - Processing text through `model.predict()` with `k=1` to get the top language prediction
-- Removing the `__label__` prefix while preserving standard FastText's uppercase output and complete GlotLID labels (for example, `__label__en` becomes `EN` and `__label__eng_Latn` becomes `eng_Latn`)
+- Removing the `__label__` prefix while preserving the model label's original casing (for example, `__label__en` becomes `en` and `__label__eng_Latn` becomes `eng_Latn`)
 - Comparing confidence scores against the threshold to determine document retention
 - Returning results as `[confidence_score, language_code]` for downstream processing
 
@@ -132,7 +132,7 @@ The language identification process adds a score field to each document batch:
 
 1. **`language` field**: Contains the FastText language identification results as a string representation of a list with two elements (for backend compatibility):
    - Element 0: The confidence score (between 0 and 1)
-   - Element 1: The standard uppercase language code or complete GlotLID label without the `__label__` prefix (for example, `EN` or `eng_Latn`)
+   - Element 1: The model label without the `__label__` prefix and with its original casing preserved (for example, `en` or `eng_Latn`)
 
 2. **Task-based processing**: Curator processes documents in batches (tasks), and results are available through the task's Pandas DataFrame:
 
diff --git a/nemo_curator/stages/text/filters/fasttext/fasttext_filters.py b/nemo_curator/stages/text/filters/fasttext/fasttext_filters.py
index 9e678adb6e..8bee989416 100644
--- a/nemo_curator/stages/text/filters/fasttext/fasttext_filters.py
+++ b/nemo_curator/stages/text/filters/fasttext/fasttext_filters.py
@@ -69,7 +69,7 @@ def __init__(self, model_path: str | None = None, min_langid_score: float = 0.3,
             msg = "Must provide a valid path to a FastText model to identify languages with this filter"
             raise ValueError(msg)
         self._model_path = model_path
-        self._lang_code = lang.casefold() if lang else None
+        self._lang_code = lang or None
         self._cutoff = min_langid_score
         self._name = "lang_id"
 
@@ -89,8 +89,6 @@ def score_document(self, text: str) -> str:
         label, score = model.predict([pp], k=1)
         score = score[0][0].item()
         lang_code = label[0][0].removeprefix("__label__")
-        if "_" not in lang_code:
-            lang_code = lang_code.upper()
 
         # Need to convert it to a string to allow backend conversions
         return str([score, lang_code])
@@ -104,9 +102,10 @@ def keep_document(self, score: float | str) -> bool:
             msg = "score must be a string convertible to list"
             raise TypeError(msg)
         if self._lang_code:
-            if "_" in self._lang_code:
-                language_matches = lang == self._lang_code
+            lang_filter = self._lang_code.casefold()
+            if "_" in lang_filter:
+                language_matches = lang == lang_filter
             else:
-                language_matches = lang.split("_", maxsplit=1)[0] == self._lang_code
+                language_matches = lang.split("_", maxsplit=1)[0] == lang_filter
             return score >= self._cutoff and language_matches
         return score >= self._cutoff
diff --git a/tests/stages/text/filters/__init__.py b/tests/stages/text/filters/__init__.py
index 4fc25d0d3c..e69de29bb2 100644
--- a/tests/stages/text/filters/__init__.py
+++ b/tests/stages/text/filters/__init__.py
@@ -1,13 +0,0 @@
-# Copyright (c) 2026, NVIDIA CORPORATION.  All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
diff --git a/tests/stages/text/filters/fasttext/__init__.py b/tests/stages/text/filters/fasttext/__init__.py
index 4fc25d0d3c..e69de29bb2 100644
--- a/tests/stages/text/filters/fasttext/__init__.py
+++ b/tests/stages/text/filters/fasttext/__init__.py
@@ -1,13 +0,0 @@
-# Copyright (c) 2026, NVIDIA CORPORATION.  All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
diff --git a/tests/stages/text/filters/fasttext/test_fasttext_filters.py b/tests/stages/text/filters/fasttext/test_fasttext_filters.py
index 48f3942b13..7803eb6da8 100644
--- a/tests/stages/text/filters/fasttext/test_fasttext_filters.py
+++ b/tests/stages/text/filters/fasttext/test_fasttext_filters.py
@@ -15,15 +15,84 @@
 from unittest.mock import Mock
 
 import numpy as np
+import pandas as pd
 import pytest
 
+from nemo_curator.stages.text.filters import DocumentFilter, ScoreFilter
 from nemo_curator.stages.text.filters.fasttext import FastTextLangId
+from nemo_curator.tasks import DocumentBatch
+
+
+class FakeQualityFilter(DocumentFilter):
+    """Emulate ``FastTextQualityFilter`` without loading a model."""
+
+    def __init__(self, alpha: float = 3, seed: int = 42):
+        super().__init__()
+        self._alpha = alpha
+        self._seed = np.random.seed(seed)  # noqa: NPY002
+
+    def load_model(self) -> None:
+        pass
+
+    def score_document(self, text: str) -> float:
+        scores = {"a": 0.00, "b": 0.25, "c": 0.50, "d": 0.75}
+        try:
+            return scores[text]
+        except KeyError:
+            msg = f"Unexpected text: {text}"
+            raise ValueError(msg) from None
+
+    def keep_document(self, score: float) -> bool:
+        return np.random.pareto(self._alpha) > 1 - score  # noqa: NPY002
+
+
+class FakeLangId(DocumentFilter):
+    """Emulate ``FastTextLangId`` without loading a model."""
+
+    def __init__(self, min_langid_score: float = 0.3):
+        super().__init__()
+        self._cutoff = min_langid_score
+
+    def load_model(self) -> None:
+        pass
+
+    def score_document(self, text: str) -> str:
+        scores = {
+            "a": [0.5, "EN"],
+            "b": [0.7, "HI"],
+            "c": [0.2, "PT"],
+            "d": [0.5, "EN"],
+        }
+        try:
+            return str(scores[text])
+        except KeyError:
+            msg = f"Unexpected text: {text}"
+            raise ValueError(msg) from None
+
+    def keep_document(self, score: float | str) -> bool:
+        if isinstance(score, str):
+            score = eval(score)  # noqa: S307
+
+        return score[0] >= self._cutoff
+
+
+def list_to_dataset(documents: list[str]) -> DocumentBatch:
+    return DocumentBatch(data=pd.DataFrame({"text": documents}), dataset_name="test_1")
+
+
+def assert_datasets_equal(expected: DocumentBatch, actual: DocumentBatch) -> None:
+    pd.testing.assert_frame_equal(
+        expected.to_pandas().reset_index(drop=True),
+        actual.to_pandas().reset_index(drop=True),
+    )
+    assert actual.dataset_name == expected.dataset_name
 
 
 @pytest.mark.parametrize(
     ("label", "expected_language"),
     [
-        ("__label__en", "EN"),
+        ("__label__en", "en"),
+        ("__label__EN", "EN"),
         ("__label__eng_Latn", "eng_Latn"),
     ],
 )
@@ -60,3 +129,21 @@ def test_keep_document_applies_score_cutoff_with_glotlid_label() -> None:
     lang_id = FastTextLangId(model_path="model.bin", min_langid_score=0.8, lang="eng")
 
     assert not lang_id.keep_document(str([0.7, "eng_Latn"]))
+
+
+def test_fake_quality_filter_pipeline() -> None:
+    dataset = list_to_dataset(["a", "b", "c", "d"])
+
+    filtered_data = ScoreFilter(FakeQualityFilter()).process(dataset)
+
+    expected_data = list_to_dataset(["b", "c", "d"])
+    assert_datasets_equal(expected_data, filtered_data)
+
+
+def test_fake_langid_filter_pipeline() -> None:
+    dataset = list_to_dataset(["a", "b", "c", "d"])
+
+    filtered_data = ScoreFilter(FakeLangId()).process(dataset)
+
+    expected_data = list_to_dataset(["a", "b", "d"])
+    assert_datasets_equal(expected_data, filtered_data)
diff --git a/tests/stages/text/modules/test_filters.py b/tests/stages/text/modules/test_filters.py
index 486b43706f..ed985cd7a0 100644
--- a/tests/stages/text/modules/test_filters.py
+++ b/tests/stages/text/modules/test_filters.py
@@ -15,7 +15,6 @@
 import os
 import re
 
-import numpy as np
 import pandas as pd
 import pytest
 
@@ -87,64 +86,17 @@ def encode(self, text: str) -> list[str]:
         return text.split()
 
 
-class FakeQualityFilter(DocumentFilter):
-    """
-    Emulates FastTextQualityFilter without a model
-    """
-
-    def __init__(self, alpha: float = 3, seed: int = 42):
-        super().__init__()
-        self._alpha = alpha
-        self._seed = np.random.seed(seed)  # noqa: NPY002
+class FakeModelFilter(DocumentFilter):
+    """Minimal model-backed filter used to test actor-stage detection."""
 
     def load_model(self) -> None:
         pass
 
     def score_document(self, text: str) -> float:
-        if text == "a":
-            return 0.00
-        elif text == "b":
-            return 0.25
-        elif text == "c":
-            return 0.50
-        elif text == "d":
-            return 0.75
-        else:
-            msg = f"Unexpected text: {text}"
-            raise ValueError(msg)
+        return float(bool(text))
 
     def keep_document(self, score: float) -> bool:
-        return np.random.pareto(self._alpha) > 1 - score  # noqa: NPY002
-
-
-class FakeLangId(DocumentFilter):
-    """
-    Emulates FastTextLangId without a model
-    """
-
-    def __init__(self, min_langid_score: float = 0.3):
-        super().__init__()
-        self._cutoff = min_langid_score
-
-    def load_model(self) -> None:
-        pass
-
-    def score_document(self, text: str) -> str:
-        if text in ["a", "d"]:
-            return str([0.5, "EN"])
-        if text == "b":
-            return str([0.7, "HI"])
-        if text == "c":
-            return str([0.2, "PT"])
-        else:
-            msg = f"Unexpected text: {text}"
-            raise ValueError(msg)
-
-    def keep_document(self, score: float | str) -> bool:
-        if isinstance(score, str):
-            score = eval(score)  # noqa: S307
-
-        return score[0] >= self._cutoff
+        return bool(score)
 
 
 def all_equal(left_dataset: DocumentBatch, right_dataset: DocumentBatch) -> bool:
@@ -451,13 +403,9 @@ def test_ray_stage_spec(self) -> None:
         assert test_filter.ray_stage_spec() == {"is_actor_stage": False}
 
         # Has load_model
-        test_filter = ScoreFilter(FakeQualityFilter(), text_field="documents")
-        assert test_filter.ray_stage_spec() == {"is_actor_stage": True}
-        test_filter = Score(FakeQualityFilter(), text_field="documents", score_field="score")
+        test_filter = ScoreFilter(FakeModelFilter(), text_field="documents")
         assert test_filter.ray_stage_spec() == {"is_actor_stage": True}
-        test_filter = ScoreFilter(FakeLangId(), text_field="documents")
-        assert test_filter.ray_stage_spec() == {"is_actor_stage": True}
-        test_filter = Score(FakeLangId(), text_field="documents", score_field="score")
+        test_filter = Score(FakeModelFilter(), text_field="documents", score_field="score")
         assert test_filter.ray_stage_spec() == {"is_actor_stage": True}
 
         # Has load_tokenizer
@@ -1288,29 +1236,3 @@ def test_line_statistics(
     ) -> None:
         line_statistics = per_extension_filter._line_statistics(content)
         assert line_statistics == expected, f"Expected {expected} but got {line_statistics}"
-
-
-class TestClassifierFilters:
-    def test_fake_quality_filter(self) -> None:
-        dataset = list_to_dataset(["a", "b", "c", "d"])
-        filters = ScoreFilter(FakeQualityFilter())
-
-        filtered_data = filters.process(dataset)
-
-        expected_data = DocumentBatch(
-            data=pd.DataFrame({"text": ["b", "c", "d"]}),
-            dataset_name="test_1",
-        )
-        assert all_equal(expected_data, filtered_data), f"Expected {expected_data} but got {filtered_data}"
-
-    def test_fake_langid_filter(self) -> None:
-        dataset = list_to_dataset(["a", "b", "c", "d"])
-        filters = ScoreFilter(FakeLangId())
-
-        filtered_data = filters.process(dataset)
-
-        expected_data = DocumentBatch(
-            data=pd.DataFrame({"text": ["a", "b", "d"]}),
-            dataset_name="test_1",
-        )
-        assert all_equal(expected_data, filtered_data), f"Expected {expected_data} but got {filtered_data}"