Why We Built This

‍

The internet runs on language. It powers conversations, fuels communities, and drives entire businesses. Yet not all language is healthy. Toxic or aggressive comments can quickly erode trust, harm users, and create reputational risks for organizations that host online communication.

‍

We built this Text Detoxification Model with Minibase to help teams moderate, rewrite, and transform harmful language into constructive, respectful communication. The model is designed to detect toxic, profane, or inflammatory text and rephrase it into language that maintains the original meaning while removing aggression, discrimination, or personal attacks.

‍

Unlike traditional moderation tools that simply block or delete content, this model repairs it. It understands context, tone, and intent, allowing it to reframe harmful messages rather than silence them. The result is a practical, real-time safeguard for digital platforms, education systems, customer service channels, and social media products that depend on healthy conversation.

‍

Key Features:

‍

• Fast and Lightweight: The model is optimized for CPU-only environments and identifies entities in milliseconds without the need for GPUs.
• Domain Adaptable: Trained across diverse sources, it performs well on technical, legal, financial, and general text, and it can be easily fine-tuned for custom taxonomies or industries.
• Privacy-Preserving: All processing can be done locally, ensuring that private messages, reports, and confidential data never leave your environment.
• Contextual Accuracy: The model understands sentence structure and entity boundaries even in complex text, handling abbreviations, dates, and multi-token names naturally.

Use Case Examples:

‍

• Finance and Compliance: Identify company names, transaction references, and regulatory entities in contracts or filings to flag risks and obligations automatically.
• Healthcare: Extract patient names, medications, dosages, and medical conditions from clinical notes or research papers while maintaining data privacy.
• Government and Intelligence: Parse field reports, public documents, or communications to detect people, organizations, and locations in real time.
• Customer Support and Operations: Highlight order numbers, customer IDs, and product names within large message logs to streamline routing and case resolution.
• Media and Research: Tag people, events, and organizations in large text archives to power advanced search tools or build knowledge graphs.

‍

Creation Journey

‍

When we set out to create this model, our goal was to design a high-accuracy, low-latency system capable of understanding complex language and extracting structured meaning from it. Named Entity Recognition may appear straightforward, but in practice it requires deep contextual understanding to work reliably across messy real-world text.

‍

Datasets, Fueling the Learning Process

‍

We began by curating a balanced dataset that included the most relevant entity categories: locations, organizations, people, and miscellaneous entities such as events, dates, or creative works. Each example was labeled carefully so that every token corresponded precisely to the correct entity. We sourced text from a range of domains including news, legal documents, technical manuals, and online conversations. By including both clean and informal language, the dataset allowed the model to perform well on a wide variety of real-world inputs.

‍

The data preparation phase was extensive. We normalized capitalization, standardized punctuation, and cleaned noisy text while preserving subtle contextual clues such as how commas or quotation marks can alter entity boundaries. We also generated synthetic examples using Minibase’s dataset builder to teach the model how to handle uncommon patterns such as multi-word names, hyphenated company titles, and nested entities.

‍

>> Want to create your own synthetic dataset?

‍

Training and Fine-Tuning

‍

Once the dataset was ready, we selected a small base model optimized for sequence labeling tasks. Using Minibase, we fine-tuned the model with a parameter-efficient training process that monitored precision, recall, and F1-score across entity types. Minibase provided automated tracking, validation, and version control, which allowed us to focus entirely on improving quality rather than managing infrastructure.

‍

After training, we performed several optimization steps to prepare the model for deployment. Quantization and layer pruning significantly reduced its size and memory requirements without affecting accuracy. The resulting artifact ran efficiently on laptops, servers, and embedded devices. During evaluation, the model demonstrated a strong ability to differentiate between entities that share names but differ in meaning, such as “Apple” the company and “apple” the fruit, or “Amazon” the organization and “Amazon” the river. These contextual distinctions were key to delivering high trust and usability in real applications.

‍

By the end of the process, we had built a high-performing model that balanced precision, speed, and adaptability. Using Minibase, what could have taken weeks of engineering and experimentation was accomplished in a single streamlined workflow.

‍

The Result

‍

The final NER model is both accurate and efficient. It provides near real-time entity recognition for documents, chat logs, and live text streams. It achieves high precision and recall across all major entity categories and maintains consistent context even in complex or lengthy passages. In comparative evaluations, it performs competitively with much larger models while being several times faster and easier to deploy.

‍

In real use, the model can process entire documents in seconds and immediately return structured, labeled data. It identifies people, organizations, and locations with near-human accuracy, allowing teams to automate data extraction that once required extensive manual review. The model performs reliably in both structured enterprise documents and informal content such as emails or social media posts.

‍

Because it is entirely self-contained, the model can run fully offline. Organizations can integrate it into their internal systems, ensuring that sensitive information never leaves secure infrastructure. For large-scale deployments, Minibase provides packaged versions ready for use in containers, APIs, and browser applications.

From start to finish, the project took less than a day to complete using Minibase. In traditional workflows, developing an NER system of this quality could take weeks of dataset engineering, GPU allocation, and DevOps configuration. The finished model shows that advanced natural language understanding is no longer limited to large cloud services or specialized research teams.

‍

The NER model delivers a practical solution for real-world information extraction. It is small, fast, and reliable, and it demonstrates how Minibase makes it possible for any organization to build purpose-built AI models that turn raw text into structured insight.

‍

>> Want to use it for yourself? You can download it here.

‍

Create your own AI models with Minibase - the possibilities for customization are endless.

‍

>> Want to build your own model? Try Minibase now.‍

‍

>> Need us to build it for you? Contact our solutions team.

‍

‍