Content Type Classifier

README.md

@@ -28,7 +28,7 @@ All of our text pipelines have great multilingual support.
 - [Heuristic Filtering](https://docs.nvidia.com/nemo-framework/user-guide/latest/datacuration/qualityfiltering.html)
 - Classifier Filtering
   - [fastText](https://docs.nvidia.com/nemo-framework/user-guide/latest/datacuration/qualityfiltering.html)
-  - GPU-Accelerated models: [Domain (English and multilingual), Quality, Safety, and Educational Content Classification](https://docs.nvidia.com/nemo-framework/user-guide/latest/datacuration/distributeddataclassification.html)
+  - GPU-Accelerated models: [Domain (English and multilingual), Quality, Safety, Educational Content, and Content Type Classification](https://docs.nvidia.com/nemo-framework/user-guide/latest/datacuration/distributeddataclassification.html)
 - **GPU-Accelerated Deduplication**
   - [Exact Deduplication](https://docs.nvidia.com/nemo-framework/user-guide/latest/datacuration/gpudeduplication.html)
   - [Fuzzy Deduplication](https://docs.nvidia.com/nemo-framework/user-guide/latest/datacuration/gpudeduplication.html) via MinHash Locality Sensitive Hashing

docs/user-guide/api/classifiers.rst

@@ -19,3 +19,6 @@ Classifiers
 
 .. autoclass:: nemo_curator.classifiers.InstructionDataGuardClassifier
     :members:
+
+.. autoclass:: nemo_curator.classifiers.ContentTypeClassifier
+    :members:

docs/user-guide/cpuvsgpu.rst

@@ -71,6 +71,7 @@ The following NeMo Curator modules are GPU based.
   * Quality Classification
   * AEGIS and Instruction-Data-Guard Safety Models
   * FineWeb Educational Content Classification
+  * Content Type Classification
 
 GPU modules store the ``DocumentDataset`` using a ``cudf`` backend instead of a ``pandas`` one.
 To read a dataset into GPU memory, one could use the following function call.

docs/user-guide/distributeddataclassification.rst

@@ -15,7 +15,7 @@ NeMo Curator provides a module to help users run inference with pre-trained mode
 This is achieved by chunking the datasets across multiple computing nodes, each equipped with multiple GPUs, to accelerate the classification task in a distributed manner.
 Since the classification of a single text document is independent of other documents within the dataset, we can distribute the workload across multiple nodes and GPUs to perform parallel processing.
 
-Domain (English and multilingual), quality, content safety, and educational content models are tasks we include as examples within our module.
+Domain (English and multilingual), quality, content safety, educational content, and content type models are tasks we include as examples within our module.
 
 Here, we summarize why each is useful for training an LLM:
 
@@ -31,6 +31,8 @@ Here, we summarize why each is useful for training an LLM:
 
 - The **FineWeb Educational Content Classifier** focuses on identifying and prioritizing educational material within datasets. This classifier is especially useful for training LLMs on specialized educational content, which can improve their performance on knowledge-intensive tasks. Models trained on high-quality educational content demonstrate enhanced capabilities on academic benchmarks such as MMLU and ARC, showcasing the classifier's impact on improving the knowledge-intensive task performance of LLMs.
 
+- The **Content Type Classifier** is designed to categorize documents into one of 11 distinct speech types based on their content. It analyzes and understands the nuances of textual information, enabling accurate classification across a diverse range of content types.
+
 -----------------------------------------
 Usage
 -----------------------------------------
@@ -232,6 +234,28 @@ For example, to create a dataset with only highly educational content (scores 4
     high_edu_dataset = result_dataset[result_dataset["fineweb-edu-score-int"] >= 4]
     high_edu_dataset.to_json("high_educational_content/")
 
+Content Type Classifier
+^^^^^^^^^^^^^^^^^^^^^^^
+
+The Content Type Classifier is used to categorize speech types based on their content. It analyzes and understands the nuances of textual information, enabling accurate classification across a diverse range of content types.
+
+Let's see how ``ContentTypeClassifier`` works in a small excerpt taken from ``examples/classifiers/content_type_example.py``:
+
+.. code-block:: python
+
+    from nemo_curator.classifiers import ContentTypeClassifier
+
+    files = get_all_files_paths_under("books_dataset/")
+    input_dataset = DocumentDataset.read_json(files, backend="cudf")
+
+    content_type_classifier = ContentTypeClassifier(filter_by=["Blogs", "News"])
+    result_dataset = content_type_classifier(dataset=input_dataset)
+
+    result_dataset.to_json("blogs_and_news/")
+
+In this example, the content type classifier is obtained directly from `Hugging Face <https://huggingface.co/nvidia/content-type-classifier-deberta>`_.
+It filters the input dataset to include only documents classified as "Blogs" or "News".
+
 -----------------------------------------
 CrossFit Integration
 -----------------------------------------

examples/classifiers/README.md

@@ -8,6 +8,7 @@ The Python scripts in this directory demonstrate how to run classification on yo
 - AEGIS Safety Models
 - Instruction-Data-Guard Model
 - FineWeb Educational Content Classifier
+- Content Type Classifier
 
 For more information about these classifiers, please see NeMo Curator's [Distributed Data Classification documentation](https://docs.nvidia.com/nemo-framework/user-guide/latest/datacuration/distributeddataclassification.html).
 

examples/classifiers/content_type_example.py

@@ -0,0 +1,65 @@
+# Copyright (c) 2024, 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.
+
+import argparse
+import time
+
+from nemo_curator.classifiers import ContentTypeClassifier
+from nemo_curator.datasets import DocumentDataset
+from nemo_curator.utils.distributed_utils import get_client
+from nemo_curator.utils.script_utils import ArgumentHelper
+
+
+def main(args):
+    global_st = time.time()
+
+    # Input can be a string or list
+    input_file_path = "/path/to/data"
+    output_file_path = "./"
+
+    client_args = ArgumentHelper.parse_client_args(args)
+    client_args["cluster_type"] = "gpu"
+    client = get_client(**client_args)
+
+    input_dataset = DocumentDataset.read_json(
+        input_file_path, backend="cudf", add_filename=True
+    )
+
+    content_type_classifier = ContentTypeClassifier(filter_by=["Blogs", "News"])
+    result_dataset = content_type_classifier(dataset=input_dataset)
+
+    result_dataset.to_json(output_path=output_file_path, write_to_filename=True)
+
+    global_et = time.time()
+    print(
+        f"Total time taken for content type classifier inference: {global_et-global_st} s",
+        flush=True,
+    )
+
+    client.close()
+
+
+def attach_args(
+    parser=argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    ),
+):
+    argumentHelper = ArgumentHelper(parser)
+    argumentHelper.add_distributed_classifier_cluster_args()
+
+    return argumentHelper.parser
+
+
+if __name__ == "__main__":
+    main(attach_args().parse_args())

nemo_curator/classifiers/__init__.py

@@ -16,6 +16,7 @@
 
 os.environ["RAPIDS_NO_INITIALIZE"] = "1"
 from .aegis import AegisClassifier, InstructionDataGuardClassifier
+from .content_type import ContentTypeClassifier
 from .domain import DomainClassifier, MultilingualDomainClassifier
 from .fineweb_edu import FineWebEduClassifier
 from .quality import QualityClassifier
@@ -27,4 +28,5 @@
     "AegisClassifier",
     "InstructionDataGuardClassifier",
     "FineWebEduClassifier",
+    "ContentTypeClassifier",
 ]

nemo_curator/classifiers/content_type.py

@@ -0,0 +1,138 @@
+# Copyright (c) 2024, 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.
+import os
+from dataclasses import dataclass
+from typing import List, Optional
+
+os.environ["RAPIDS_NO_INITIALIZE"] = "1"
+from crossfit.backend.torch.hf.model import HFModel
+from transformers import AutoConfig, AutoTokenizer
+
+from nemo_curator.classifiers.base import (
+    DistributedDataClassifier,
+    HFDeberta,
+    _get_suggest_memory_for_classifier,
+    _run_classifier_helper,
+)
+from nemo_curator.datasets import DocumentDataset
+
+CONTENT_TYPE_IDENTIFIER = "nvidia/content-type-classifier-deberta"
+
+
+@dataclass
+class ContentTypeModelConfig:
+    model: str = "microsoft/deberta-v3-base"
+    fc_dropout: float = 0.2
+    max_len: int = 1024
+
+
+class ContentTypeModel(HFModel):
+    def __init__(
+        self,
+        config: ContentTypeModelConfig,
+        autocast: bool = False,
+        max_mem_gb: Optional[int] = None,
+    ):
+        self.config = config
+        self.autocast = autocast
+        if max_mem_gb is None:
+            max_mem_gb = _get_suggest_memory_for_classifier()
+
+        super().__init__(self.config.model, max_mem_gb=max_mem_gb)
+
+    def load_model(self, device: str = "cuda"):
+        model = HFDeberta.from_pretrained(CONTENT_TYPE_IDENTIFIER)
+        model.set_autocast(self.autocast)
+        model = model.to(device)
+        return model.eval()
+
+    def load_tokenizer(self):
+        return AutoTokenizer.from_pretrained(CONTENT_TYPE_IDENTIFIER)
+
+    def load_config(self):
+        return AutoConfig.from_pretrained(CONTENT_TYPE_IDENTIFIER)
+
+
+class ContentTypeClassifier(DistributedDataClassifier):
+    """
+    ContentTypeClassifier is a text classification model designed to categorize documents into one of 11 distinct speech types based on their content.
+    It analyzes and understands the nuances of textual information, enabling accurate classification across a diverse range of content types.
+    The pretrained model used by this class can be found on Hugging Face here: https://huggingface.co/nvidia/content-type-classifier-deberta.
+    This classifier is optimized for running on multi-node, multi-GPU setups to enable fast and efficient inference on large datasets.
+
+    Attributes:
+        filter_by (list[str], optional): The classes to filter the dataset by.
+                                         If None, all classes will be included. Defaults to None.
+        batch_size (int): The number of samples per batch for inference. Defaults to 256.
+        text_field (str): The field in the dataset that should be classified.
+        pred_column (str): The column name where predictions will be stored. Defaults to "content_pred".
+        prob_column (str, optional): The column name where prediction probabilities will be stored. Defaults to None.
+        max_chars (int): The maximum number of characters in each document to consider for classification. Defaults to 5000.
+        device_type (str): The type of device to use for inference, either "cuda" or "cpu". Defaults to "cuda".
+        autocast (bool): Whether to use mixed precision for faster inference. Defaults to True.
+        max_mem_gb (int, optional): The maximum amount of memory in GB to allocate for the model. If None,
+                                      it defaults to the available GPU memory minus 4 GB.
+
+    """
+
+    def __init__(
+        self,
+        filter_by: Optional[List[str]] = None,
+        batch_size: int = 256,
+        text_field: str = "text",
+        pred_column: str = "content_pred",
+        prob_column: Optional[str] = None,
+        max_chars: int = 5000,
+        device_type: str = "cuda",
+        autocast: bool = True,
+        max_mem_gb: Optional[int] = None,
+    ):
+        config = AutoConfig.from_pretrained(CONTENT_TYPE_IDENTIFIER)
+
+        self.text_field = text_field
+        self.prob_column = prob_column
+        self.labels = list(config.label2id.keys())
+        self.labels.sort(key=lambda x: config.label2id[x])
+        self.out_dim = len(self.labels)
+
+        model = ContentTypeModel(
+            config=ContentTypeModelConfig, autocast=autocast, max_mem_gb=max_mem_gb
+        )
+
+        super().__init__(
+            model=model,
+            labels=self.labels,
+            filter_by=filter_by,
+            batch_size=batch_size,
+            out_dim=self.out_dim,
+            pred_column=pred_column,
+            max_chars=max_chars,
+            device_type=device_type,
+            autocast=autocast,
+        )
+
+    def _run_classifier(self, dataset: DocumentDataset) -> DocumentDataset:
+        print("Starting content type classifier inference", flush=True)
+        df = dataset.df
+        df = _run_classifier_helper(
+            df=df,
+            model=self.model,
+            labels=self.labels,
+            max_chars=self.max_chars,
+            batch_size=self.batch_size,
+            label_col=self.pred_column,
+            text_field=self.text_field,
+            prob_col=self.prob_column,
+        )
+        return DocumentDataset(df)

nemo_curator/classifiers/domain.py

@@ -148,7 +148,7 @@ class DomainClassifier(_DomainClassifier):
     """
     DomainClassifier is a specialized classifier designed for English text domain classification tasks,
     utilizing the NVIDIA Domain Classifier (https://huggingface.co/nvidia/domain-classifier) model.
-    This class is optimized for running on multi-node, multi-GPU setups to enable fast and efficient inference on large datasets.
+    This classifier is optimized for running on multi-node, multi-GPU setups to enable fast and efficient inference on large datasets.
 
     Attributes:
         filter_by (list[str], optional): The classes to filter the dataset by.
@@ -196,7 +196,7 @@ class MultilingualDomainClassifier(_DomainClassifier):
     MultilingualDomainClassifier is a specialized classifier designed for domain classification tasks,
     utilizing the NVIDIA Multilingual Domain Classifier (https://huggingface.co/nvidia/multilingual-domain-classifier) model.
     It supports domain classification across 52 languages.
-    This class is optimized for running on multi-node, multi-GPU setups to enable fast and efficient inference on large datasets.
+    This classifier is optimized for running on multi-node, multi-GPU setups to enable fast and efficient inference on large datasets.
 
     Attributes:
         filter_by (list[str], optional): The classes to filter the dataset by.

nemo_curator/classifiers/fineweb_edu.py

@@ -69,10 +69,9 @@ def load_config(self):
 
 class FineWebEduClassifier(DistributedDataClassifier):
     """
-    FineWebEduClassifier is a specialized classifier designed for educational content assessment, utilizing the
-    Hugging Face FineWeb EDU Classifier model (https://huggingface.co/HuggingFaceFW/fineweb-edu-classifier).
-    This class is optimized for running on multi-node, multi-GPU setups to enable fast and efficient inference
-    on large text datasets.
+    FineWebEduClassifier is a specialized classifier designed for educational content assessment,
+    utilizing the Hugging Face FineWeb EDU Classifier model (https://huggingface.co/HuggingFaceFW/fineweb-edu-classifier).
+    This classifier is optimized for running on multi-node, multi-GPU setups to enable fast and efficient inference on large text datasets.
 
     Attributes:
         batch_size (int): The number of samples per batch for inference. Defaults to 256.

nemo_curator/classifiers/quality.py

@@ -65,9 +65,9 @@ def load_config(self):
 
 class QualityClassifier(DistributedDataClassifier):
     """
-    QualityClassifier is a specialized classifier designed for quality assessment tasks, utilizing the
-    NVIDIA Quality Classifier model (https://huggingface.co/nvidia/quality-classifier-deberta). This class is
-    optimized for running on multi-node, multi-GPU setups to enable fast and efficient inference on large datasets.
+    QualityClassifier is a specialized classifier designed for quality assessment tasks,
+    utilizing the NVIDIA Quality Classifier model (https://huggingface.co/nvidia/quality-classifier-deberta).
+    This classifier is optimized for running on multi-node, multi-GPU setups to enable fast and efficient inference on large datasets.
 
     Attributes:
         filter_by (list[str], optional): The classes to filter the dataset by. If None, all classes will be included. Defaults to None.

nemo_curator/scripts/classifiers/README.md

@@ -8,6 +8,7 @@ The Python scripts in this directory demonstrate how to run classification on yo
 - AEGIS Safety Models
 - Instruction-Data-Guard Model
 - FineWeb Educational Content Classifier
+- Content Type Classifier
 
 For more information about these classifiers, please see NeMo Curator's [Distributed Data Classification documentation](https://docs.nvidia.com/nemo-framework/user-guide/latest/datacuration/distributeddataclassification.html).
 
@@ -136,3 +137,22 @@ fineweb_edu_classifier_inference \
 ```
 
 Additional arguments may be added for customizing a Dask cluster and client. Run `fineweb_edu_classifier_inference --help` for more information.
+
+#### Content type classifier inference
+
+```bash
+# same as `python content_type_classifier_inference.py`
+content_type_classifier_inference \
+    --input-data-dir /path/to/data/directory \
+    --output-data-dir /path/to/output/directory \
+    --input-file-type "jsonl" \
+    --input-file-extension "jsonl" \
+    --output-file-type "jsonl" \
+    --input-text-field "text" \
+    --batch-size 64 \
+    --autocast \
+    --max-chars 5000 \
+    --device "gpu"
+```
+
+Additional arguments may be added for customizing a Dask cluster and client. Run `content_type_classifier_inference --help` for more information.