NeMo Curator Classifier Models: How Domain and Quality Classification Creates High-Quality Data Blends

Why Data Classification Matters for LLM Training

Building a high-quality LLM requires more than collecting massive amounts of text. Raw web crawl data contains enormous variation in topic coverage, writing quality, and domain relevance. Without classification, training datasets end up imbalanced — overrepresenting some domains while underrepresenting others, mixing high-quality academic content with low-quality spam.

NeMo Curator provides GPU-accelerated classifier models that categorize text by domain and quality, enabling teams to create balanced, high-quality data blends specifically tuned for their model's target use cases.

The Value Proposition of NeMo Curator Classification

Accelerated Inference

NeMo Curator leverages RAPIDS, NVIDIA's GPU-accelerated data science toolkit, for distributed data classification. Intelligent batching maximizes GPU throughput and reduces latency when classifying millions of text samples. What would take days on CPU-based systems completes in hours on GPU infrastructure.

flowchart LR
    CORPUS[("Pre-training corpus<br/>trillions of tokens")]
    FILTER["Quality filter and<br/>dedupe"]
    TOK["BPE tokenizer"]
    SHARD["Shard plus<br/>data parallel"]
    GPU{"GPU cluster<br/>FSDP or DeepSpeed"}
    CKPT[("Checkpoints<br/>every N steps")]
    LOSS["Loss curve plus<br/>eval gates"]
    SFT["SFT phase"]
    DPO["DPO or RLHF"]
    BASE([Base model])
    INSTR([Instruct model])
    CORPUS --> FILTER --> TOK --> SHARD --> GPU
    GPU --> CKPT --> LOSS
    LOSS --> BASE --> SFT --> DPO --> INSTR
    style GPU fill:#4f46e5,stroke:#4338ca,color:#fff
    style LOSS fill:#f59e0b,stroke:#d97706,color:#1f2937
    style INSTR fill:#059669,stroke:#047857,color:#fff

Seamless Scalability

The classification system handles terabyte-scale datasets without performance bottlenecks. This scalability is essential for LLM data pipelines where datasets routinely exceed hundreds of gigabytes of text.

Parallelized Processing

Classification workloads run in parallel across multiple GPUs, achieving near-linear speedup. A dataset that takes 24 hours on a single GPU processes in approximately 3 hours on eight GPUs.

Efficient Resource Usage

NeMo Curator's classifier models are lightweight, open-source models released under the Apache 2.0 license. They process massive datasets with reduced hardware requirements compared to using full LLMs for classification.

Extensible Model Support

Two core classifier models are currently available, with a roadmap to expand support for additional categories including topic relevance, style classification, and safety filters.

Domain Classifier

The Domain Classifier categorizes text into specific knowledge or topic areas. With over 250,000 downloads, it is NeMo Curator's most widely adopted model.

Supported Classes

The model classifies text into 26 domain categories. The top 10 most common classifications are:

1. Finance — Banking, investing, economics, and financial markets
2. Health — Medical, wellness, pharmaceutical, and healthcare content
3. Business and Industrial — Corporate, manufacturing, and industrial topics
4. Science — Physics, chemistry, biology, and research content
5. Law and Government — Legal, regulatory, and government policy content
6. Internet and Telecom — Digital services, networking, and telecommunications
7. Jobs and Education — Employment, career, and educational content
8. News — Current events, journalism, and media coverage
9. Computers and Electronics — Technology, hardware, and software content
10. Shopping — E-commerce, retail, and consumer product content

Training Data

The Domain Classifier was trained on 1 million Common Crawl samples and 500,000 Wikipedia articles. This combination ensures broad coverage across knowledge domains while maintaining classification accuracy on both web-crawled and encyclopedic content.

Use Cases

Domain classification enables teams to create balanced training data blends. If your model needs strong performance in healthcare and finance, you can filter for those domains and ensure proportional representation. Without domain classification, web-crawled datasets typically overrepresent shopping and news content while underrepresenting science and legal content.

Quality Classifier

The Quality Classifier evaluates document quality using linguistic and informational metrics. With over 12,000 downloads, it serves as the quality gate in data curation pipelines.

Quality Labels

Each document receives one of three quality ratings:

• High — Well-written, informative, and factually grounded content suitable for direct use in training
• Medium — Acceptable quality with some issues; may need additional filtering or editing
• Low — Poorly written, uninformative, or spam content that should be excluded from training data

Evaluation Criteria

The Quality Classifier was trained on human annotations evaluating multiple factors:

• Writing quality: Grammar, clarity, and structural coherence
• Informativeness: Depth and usefulness of the information presented
• Factual grounding: Whether claims are supported by evidence
• Relevance: Whether the content provides value for its apparent purpose
• Readability: Ease of comprehension for the target audience

Use Cases

Quality classification is the most impactful single step in data curation. Removing low-quality content from training data consistently improves model performance across benchmarks. The Quality Classifier automates what would otherwise require human reviewers, scaling quality assessment from thousands to billions of documents.

Building Data Blends

The real power of NeMo Curator's classifiers emerges when Domain and Quality classification work together. A typical workflow:

1. Classify by domain to understand the topic distribution of your raw dataset
2. Classify by quality to identify the proportion of high, medium, and low quality content in each domain
3. Filter by removing all low-quality content and optionally removing medium-quality content
4. Balance the remaining data across domains according to your model's target use case
5. Blend the balanced, filtered data into a final training dataset

This pipeline ensures that every sample in your training data is both topically relevant and meets quality standards — two properties that are essential for training reliable LLMs.

Frequently Asked Questions

What is NeMo Curator's Domain Classifier?

NeMo Curator's Domain Classifier is a GPU-accelerated model that categorizes text documents into 26 knowledge domains (Finance, Health, Science, Law, etc.). Trained on 1 million Common Crawl samples and 500,000 Wikipedia articles, it processes terabyte-scale datasets using NVIDIA RAPIDS for distributed classification. It helps teams create balanced training data blends for LLM development.

How does the Quality Classifier evaluate documents?

The Quality Classifier assigns each document a High, Medium, or Low quality rating based on writing quality, informativeness, factual grounding, relevance, and readability. It was trained on human-annotated data where reviewers evaluated these factors. The classifier automates quality assessment at scale, enabling teams to filter out low-quality content from datasets containing billions of documents.

Can NeMo Curator classifiers run on multiple GPUs?

Yes. NeMo Curator classifiers leverage NVIDIA RAPIDS for distributed processing across multiple GPUs. Classification workloads achieve near-linear speedup with additional GPUs, meaning a dataset that takes 24 hours on one GPU processes in approximately 3 hours on eight GPUs. This scalability is essential for terabyte-scale LLM data pipelines.

What is a data blend in LLM training?

A data blend is a curated mix of training data balanced across domains and quality levels. Rather than training on raw web crawl data (which overrepresents some topics and includes low-quality content), teams use classifiers to filter and balance data according to their model's target use case. Well-designed data blends consistently outperform larger but unbalanced datasets.

Are the NeMo Curator classifiers open source?

Yes. Both the Domain Classifier and Quality Classifier are released under the Apache 2.0 license. They are lightweight models optimized for efficient classification, reducing hardware requirements compared to using full-size LLMs for the same task. The models are available on Hugging Face and integrate directly with the NeMo Curator pipeline.