What is Document Classification?

Document Classification is a process where a machine learning model categorizes documents into predefined classes. This capability allows for more streamlined processing of large volumes of documents by automatically organizing them according to their content. At Taτsu, we have integrated this functionality to handle a variety of document types, enhancing both the efficiency and accuracy of document processing on our platform.

Supported Document Types

With the integration of Document Classification, our platform now supports the classification of sixteen distinct document types, each crucial in various professional settings:

• Advertisement
• Budget
• Email
• File Folder
• Form
• Handwritten
• Invoice
• Letter
• Memo
• News Article
• Presentation
• Questionnaire
• Resume
• Scientific Publication
• Scientific Report
• Specifications

These additions ensure that our Document Understanding platform can serve a broader range of industries and functions, from academic environments to corporate settings, providing versatile document management solutions.

The Path to Enhanced Accuracy

It’s important to note that as with any advanced AI functionality newly released in a live environment, initial imperfections are anticipated. The purpose of launching the Document Classification function in its current state is to continually refine it through real-world application and feedback. This process involves miners working on synthetic data generated by validators, which helps in enhancing the model’s accuracy and reliability over time.

Improved Platform Security with New Registration System

To complement the rollout of new functionalities and to safeguard the integrity of our platform, we have introduced a mandatory registration system for users. This system not only helps prevent potential abuse but also enhances the overall security and user experience of the platform. By requiring users to register and verify their email, we ensure that only legitimate users access and utilize the platform’s capabilities.

Looking Forward

As Taτsu strides forward, integrating advanced functionalities like Document Classification marks a pivotal development in our journey. This integration is just one of the many steps we are taking to solidify our position in the decentralized AI space. We are excited about the future possibilities and are committed to continuous improvement and expansion of our subnet’s capabilities.

We are excited to announce a major update to our Document Understanding subnet code — Document Classification! This new functionality enhances our capabilities, allowing us to process documents with greater efficiency and precision.

Powered by the DONUT Model

To implement document classification, we have integrated the DONUT model, leveraging its robust capabilities to accurately classify documents across various categories. This addition complements our existing functionalities and ensures a more comprehensive approach to document processing.

Alongside document classification, we have also enhanced our validation mechanism. This updated system not only supports document classification but also improves validation for our Checkbox Detector, ensuring higher accuracy and reliability across all document processing tasks.

To accommodate this new functionality, we have updated our miner model to handle both document classification and checkbox detection seamlessly. With this improvement, miners can now execute two types of tasks: Document Classification and Checkbox Detection. This dual-task capability enhances the efficiency of our network, allowing miners to contribute more effectively to document understanding.

Live on Bittensor Testnet

We are thrilled to share that these latest changes have been pushed to our GitHub repository and are now live on the Bittensor testnet! This marks a significant step forward in improving document processing capabilities within our subnet.

What’s Next?

Our next milestone is integrating this functionality into our Document Understanding platform. We anticipate rolling out this update in the next week, further expanding the accessibility and usability of our document classification and checkbox detection system.

The launch of our Document Understanding platform marks a pivotal moment in our journey towards democratizing decentralized document understanding within the Bittensor network. This significant development not only showcases our commitment to innovation but also enhances our strategic positioning in the decentralized AI sector.

The Significance of the Platform Launch

The Document Understanding platform represents more than just technological advancement; it signifies a strategic expansion of our ecosystem. By providing a user-friendly gateway for regular users to access and interact with our subnet, we’re not only broadening the practical applications of our technology but also enhancing user engagement and adoption.

This platform serves as a critical component in demonstrating the real-world utility of our Document Understanding subnet, which is essential for our upcoming re-registration on the Bittensor mainnet.

How the Platform Works

The platform operates through a sophisticated Gateway API, which bridges the gap between our frontend interface and the backend processes handled by the Validators in the Bittensor network. Here’s a breakdown of how it functions:

• Frontend: Users interact with the system via a clean and intuitive interface where they can upload documents/images. The current live function, Checkbox Detection, allows for the identification and classification of checkboxes and their associated texts from uploaded images.
• Gateway API: This component acts as the middleman. It manages task queues and statuses, ensuring seamless communication between the frontend and our Validator.
• Validator and Miner Interaction: Tasks processed by the Gateway API are queued for the Validator, which then coordinates with Miners on the network to analyze and process the data. Results are subsequently stored securely and relayed back to the user through the Gateway API.
• Results Delivery: Once processed, results are fetched by the frontend from the Gateway API, allowing users to view and interact with the data derived from their documents.

Initially, our platform is connected to the Bittensor testnet, but the ultimate goal is to transition to the mainnet once we successfully re-register the Document Understanding subnet.

Future Developments and Capabilities

While the platform currently supports Checkbox Detection, we have an ambitious roadmap to expand its functionalities. Upcoming features include Advanced OCR Engine, Document Classification, Entity Detection, Highlighted and Encircled Text Detection, and JSON Data Structuring. These enhancements will progressively roll out as we continue to refine our subnet’s capabilities.

We invite everyone to explore the platform and experience firsthand the benefits it brings. Stay tuned for more updates as we continue to push the boundaries of what’s possible in the decentralized AI sector.

At Taτsu, we are excited to unveil the mechanisms behind our Frontend and Gateway API, integral components of our Document Understanding subnet. These tools are designed to enhance user experience and streamline the interaction between our users, the system, and the underlying blockchain technology. Here’s a step-by-step guide on how each component works and interacts within our ecosystem.

1. Frontend: The User Interaction Point

The Frontend serves as the primary interface for user interaction. It is designed for simplicity and efficiency, allowing users to upload documents or images with ease. Once uploaded, these files are then sent to our system for processing. The Frontend is built to be intuitive, ensuring that users can easily navigate and use the service without needing technical expertise.

2. Gateway API: The Efficient Middleman

To bridge the gap between our Frontend and the Validator, we have developed the Gateway API. This component plays a crucial role in managing the flow of data through the system. Here’s what the Gateway API is responsible for:

• Task Reception: It receives tasks uploaded by users via the Frontend.
• Task Queuing: It stores these tasks in a Redis database, organizing them into a manageable queue. This helps keep track of each task’s status, such as “queued,” “processing,” “completed,” or “error.”
• Task Management: It ensures that all tasks are adequately queued and ready for processing by the Validator.

3. Validator: The Core Task Processor

The Validator is at the heart of our processing power. It retrieves tasks from the Redis queue, processes them, and returns the results. Here’s how it accomplishes this:

• Task Processing: Upon retrieving a task, the Validator updates its status to “processing” and begins the task.
• Interaction with Miners: The Validator sends the task to the Miner for detailed processing.
• Result Management: After the Miner completes the task, the Validator stores the results in a secure cloud Bucket and updates the task status.

4. Gateway API: Facilitating Result Retrieval

Once the Validator completes the processing, the Gateway API retrieves the results from the Bucket and prepares them for user retrieval.

• Result Retrieval: The API fetches the processed results from the storage Bucket.
• Result Transmission: It then transmits these results back to the Frontend.

5. Frontend: Displaying Results to Users

The final step in the process involves the Frontend, which receives the processed results from the Gateway API and displays them to the user. This seamless display allows users to quickly view and interact with the output of their uploaded tasks.

Our setup with the Frontend, Gateway API, and Validator is designed to offer a seamless and efficient experience from start to finish. By integrating these components with our underlying Bittensor technology, we provide a robust system that not only meets the needs of users but also leverages the power of decentralized networks for data processing.

We are proud to highlight a significant evolution within Taτsu: the launch of our own Taτsu Validator. This development marks a critical shift, positioning Taτsu not just as a subnet owner but as an integral player in the Bittensor ecosystem. Here’s a deep dive into why this move is transformative for Taτsu and its community.

The Significance of Running Our Own Validator

The Taτsu Validator is not merely a technical upgrade; it’s a strategic enhancement that empowers our Document Understanding Subnet and the Taτsu community. By establishing our own validator, we are taking a significant step towards greater autonomy and direct influence over our subnet’s access. This advancement grants us full access to the functionalities of our subnet, enabling optimal management and performance that align with our commitment to decentralization.

With the Taτsu Validator operational, we can now directly validate our Document Understanding Subnet. This capability is crucial for:

• Direct Control and Management: By operating our own validator, we gain full access to our subnet’s functionalities, allowing us to offer API access and develop front-end interfaces. This capability enhances our ability to manage and expand the utility of our subnet effectively.
• Enhanced Community Participation: Community members can now delegate their TAO to our validator, engaging more directly with our growth and benefiting from competitive yields similar to those offered by other major validators.
• Strategic Treasury Usage: Our treasury is now utilized more strategically. Initially delegated with another validator to indirectly support them, our treasury will now be directly utilized to bolster our own ecosystem, enhancing our role and strengthening the decentralized network.

Advantages for TATSU and TAO Community

By delegating to Taτsu Validator, community members support the stability and growth of Taτsu’s innovative ecosystem while earning potential yields. This engagement is vital for maintaining a healthy, decentralized network and for the success of our initiatives. Benefits include:

• Zero Take Percentage: Initially, our validator will operate with a 0% take rate, ensuring that all emissions earned are passed directly to our delegators.
• Supporting Innovation: Delegating to Taτsu Validator directly supports the development and deployment of new technologies and applications within our ecosystem, driving both growth and innovation.

Expanding Influence Across Bittensor

Our ambitions with the Taτsu Validator extend beyond our own subnet. We plan to validate other subnets within the Bittensor network, leveraging the Child/Parent Hotkey system. This approach will maximize our operational efficiency and emissions impact without the need for direct infrastructure investments on each subnet.

Future Plans and Community Engagement

As we move forward with the Taτsu Validator, we invite everyone in the Taτsu and broader TAO community to join us in this exciting journey. Delegating to our validator not only supports Taτsu but also contributes to a larger vision of a decentralization.

We will continue to provide updates on our progress, upcoming features, and additional benefits for our supporters. Join us as we embrace this new chapter in Taτsu’s history, forging paths and setting benchmarks within the Bittensor ecosystem. Together, we’re not just growing; we’re transforming the landscape of decentralized technologies.

We are excited to announce the successful registration of the Document Understanding Subnet on the Bittensor network as subnet number 54. This marks a significant milestone for Taτsu, positioning us at the forefront of decentralized document processing technologies. Our latest development harnesses a custom-engineered architecture to deliver unparalleled efficiency and accuracy in analyzing and processing document data.

Cutting-Edge Technology

The core of the Document Understanding Subnet is our YOLO Checkbox Detector, built on the robust YOLOv8-large model. This advanced detector is specifically optimized for precise checkbox detection across various document types, setting new industry standards in precision. Trained on a unique dataset of over 10,000 document images, the model includes both scanned and standard formats, ensuring comprehensive coverage and adaptability.

To validate our model’s efficacy, we conducted extensive testing with a diverse set of 300 images. This rigorous benchmarking has shown that our subnet’s checkbox detection capabilities surpass leading centralized solutions, such as Azure Form Recognizer and GPT-4 Vision, with an impressive F1-Score of 0.88. Our YOLO Checkbox Detector exemplifies the potential of decentralized AI by offering unmatched accuracy and reliability for document processing tasks within the Bittensor network.

Comprehensive Testing and Validation

Prior to our mainnet launch, the subnet underwent a series of detailed local and testnet evaluations to ensure seamless functionality. Miners on the Bittensor network utilized our tools — combining the YOLOv8 model and PyTesseract OCR engine — to execute precise document processing tasks. Validators played a crucial role, assessing the accuracy and timeliness of the processed data against pre-established ground truths.

Miner and Validator Involvement

The Document Understanding Subnet not only streamlines the process of document analysis but also encourages active participation and improvement from the network’s miners. Miners are incentivized to enhance the accuracy and efficiency of their models, directly influencing their rewards and fostering a competitive yet collaborative environment.

Validators ensure the integrity and accuracy of the subnet’s operations by employing a robust validation mechanism. This involves comparing miners’ outputs with a database of annotated images, using detailed accuracy and timing metrics to score performance.

Future Developments

Looking ahead, the Document Understanding Subnet will continue to evolve. We plan to introduce additional functionalities, expand our dataset, and refine our models based on community feedback and technological advancements. Our commitment to innovation and excellence drives continuous improvement, aiming to broaden the subnet’s applications and effectiveness.

The launch of the Document Understanding Subnet on the Bittensor network represents a pivotal advancement in the field of decentralized document processing. We are proud to contribute to the TAO ecosystem with a tool that not only enhances document analysis capabilities but also embodies the principles of transparency, collaboration, and innovation.

At Taτsu, we are excited to unveil a significant advancement in our technology stack — the Document Understanding Subnet. This innovative subnet is engineered to revolutionize the way we handle and process document data by implementing a sophisticated text extractor system.

How the Document Understanding Subnet Works?

Our approach to document processing is methodical and designed to handle documents efficiently and accurately through a multi-step process:

1. Input Handling: The subnet accepts documents in PNG format, a common and widely used file type, ensuring accessibility and ease of integration for various users and applications.
2. Detection with YOLOv8: The system uses the cutting-edge YOLOv8 model to meticulously scan each document, detecting checked boxes accurately. This crucial step ensures that all pertinent areas containing checkboxes are identified, setting the stage for precise text extraction.
3. Text Extraction with OCR: Following the detection of checkboxes, the subnet employs an OCR (Optical Character Recognition) engine, specifically tailored to extract text lines from the document with high accuracy. This OCR process captures the textual content related to the checkboxes comprehensively.
4. Intelligent Postprocessing: After text extraction, the system’s postprocessor analyzes the regions around the detected checkboxes. It then extracts and accurately associates the corresponding text with each checkbox. This intelligent post processing ensures that the extracted text is accurate and significantly reduces the errors commonly encountered in traditional systems.

Creating a Robust Dataset

To ensure the effectiveness and adaptability of our subnet, we are compiling a substantial dataset comprising 10,000 images. Each document within this dataset is meticulously annotated to establish clear ground truths for checked boxes and their corresponding text. This extensive training will enable our subnet to learn from a diverse array of document formats and styles, significantly enhancing its accuracy and versatility.

Mining

It’s essential to highlight how mining operations will be conducted within this innovative framework. Mining is structured to utilize the collective expertise and hardware of the Bittensor network’s miners to perform crucial document processing tasks that lie at the heart of our text extraction system.

Components and Tasks for Mining

1. YOLOv8 Model: The YOLOv8 model is utilized by miners to detect checkboxes in documents accurately. This model’s primary function is to ensure that all relevant checkboxes within a document are identified before any text extraction occurs.
2. Open-Source OCR Engine: Following checkbox detection, the document processing task moves to text extraction, performed using Tesseract. This powerful OCR tool is responsible for extracting text from documents precisely after checkboxes have been identified.
3. Postprocessor: This component enhances the accuracy of the system by analyzing the areas around detected checkboxes. It precisely extracts and correctly associates text, ensuring that the system captures and processes the necessary information with minimal errors.

Miners follow the specified process using these tools to process document images. They generate predictions based on their analyses, which are then sent to validators within the network for verification.

Rewards System

Miners are incentivized based on the following criteria:

1. Accuracy of the System: Rewards are given for the precision with which the model detects checkboxes and the OCR extracts the text, aiming for high accuracy in both aspects.
2. Time Taken by the System: Efficiency is also rewarded; faster processing times lead to higher rewards, encouraging miners to optimize the performance of their tools.

Incentives for Continuous Improvement

To foster a culture of continuous enhancement, miners are encouraged to improve the accuracy of both the YOLOv8 model and the OCR capabilities. Miners are permitted to collect their own data, annotate it, train new models, and replace the initial tools with their custom-developed solutions. This ongoing improvement cycle not only boosts the effectiveness of the subnet but also rewards the most dedicated and innovative miners who contribute significantly to the advancement of our system’s capabilities.

Validating

Our Document Understanding Subnet incorporates a robust validation process crucial for ensuring the efficiency and accuracy of the mining operations within the Bittensor network. This section explains the pivotal role validators play and the specific procedures they follow to uphold the integrity of our subnet’s operations.

Validation Components and Procedures

Validation is a multi-step process that ensures the accuracy and reliability of the document processing tasks performed by miners:

1. Database of Annotated Images: We maintain a comprehensive database filled with annotated images that include their corresponding ground truths. This serves as a foundation for all validation tasks.
2. Document Assignment and Evaluation: Validators randomly select document images from this database to assign to miners. Once the document is processed, validators assess the predictions made by miners, focusing on both the accuracy of the content extracted and the time efficiency of the process.

Time Decision Criteria

We set clear metrics to evaluate the time performance of the miners:

Standard Time Calculation (Tn): Experiments are conducted to establish a standard processing time (Tn) that reflects expected performance on typical machines.

Time Score Calculation: The formula used to calculate the time score is:

Time Score = Tn / (Tn + Tt + ϵ)

• Here, Tt represents the time taken by the miner, and ϵ is a small constant used to fine-tune the scoring process.

Accuracy Calculation

Accuracy evaluations are detailed and rigorous:

For Individual Box-Text Pairs:

• Checked Box Score (CBS): Scores are assigned based on the overlap between the predicted and actual checked boxes. An overlap greater than 0.95 results in a perfect score of 1, decreasing gradually for overlaps between 0.7 and 0.95. Overlaps under 0.7 receive a score of 0.
• Text Similarity Score (TS): This measures the similarity between the detected text and the actual text. A perfect match scores a 1, with scores decreasing for lesser degrees of similarity.

Individual Pair Score = (CBS + TS) / 2

For All Box-Text Pairs:

Accuracy Score = (score-1 + score-2 + … + score-n) / n

Final Score Determination

The final score for each miner is derived from a weighted average of the Time and Accuracy Scores:

Final Score = 0.3 * Time Score + 0.7 * Accuracy Score

• We adjust the weights of these scores to emphasize their relative importance as needed.

This comprehensive validation framework ensures that our subnet operates not only efficiently but with a high standard of data accuracy, allowing the Document Understanding Subnet to function reliably within the dynamic environment of the Bittensor network.

Taτsu is proud to announce the second registration of our Human Identity Verification Protocol 2.0 subnet code on the Bittensor network as Subnet 36. This milestone represents a significant step forward in our ongoing commitment to innovation and utility within the decentralized AI landscape. As we venture again into the Bittensor ecosystem, we bring with us an enhanced version of our subnet code, designed to be more robust and functional than before.

What’s new with Tatsu Identity 2.0?

The updated Taτsu Identity 2.0 subnet code has undergone substantial improvements, most notably in the expansion of human verification parameters. The original subnet utilized 9 parameters; the new version expands this to 22, significantly broadening the scope and accuracy of identity verification processes. This increase not only enhances the precision of the verifications but also aligns more closely with the complex needs of modern digital interactions, where multifaceted identity confirmation is crucial.

Community Engagement and Network Integration

With the registration of Taτsu Identity 2.0, we aim to garner more traction within the Bittensor community, particularly among miners and validators. The expanded capabilities of our subnet are designed to encourage greater participation and support from these key players, ensuring that our subnet remains competitive and relevant. The improvements in our subnet’s architecture are expected to attract more engagement, fostering a stronger presence within the network.

Additionally, we are pleased to announce that our Discord channel has returned to the mainnet section of the official Bittensor Discord, having moved from the testnet section. This move signifies our restored presence on the mainnet and provides an accessible platform for both new and existing community members to connect, discuss, and receive updates directly from our team.

Looking Ahead: Document Understanding Subnet

Even as we celebrate this registration, our team is already hard at work on a new code: the Document Understanding subnet. This new development is aimed at revolutionizing how documents are processed and understood within the Bittensor network, leveraging advanced AI to extract and interpret information with unprecedented accuracy and efficiency.

We remain dedicated to pushing the boundaries of what is possible within decentralized AI and are enthusiastic about the potential impacts of our upcoming Document Understanding subnet.

The past few weeks at Taτsu have been nothing short of dynamic, marking both a significant achievement and a learning curve. We reached a pivotal milestone by registering the Taτsu Identity Protocol as Subnet 38 on the Bittensor network, officially becoming part of the mainnet. However, our journey on the mainnet was brief, lasting almost two weeks before we faced deregistration. This turn of events highlights the need for ongoing improvements to our subnet code to meet the network’s demanding performance standards.

Deregistration Insights

The Bittensor network’s competitive environment necessitates outstanding performance, which we found challenging during our initial phase. Like many subnets before us, Taτsu faced deregistration early on. This is a common hurdle, not a setback, reflecting the rigorous and dynamic nature of the network. It calls for a return to the drawing board, prompting us to refine and enhance our technology. Our team is committed to developing a more robust and competitive codebase for our subsequent re-entry into the network.

Future Plans for the Subnet

While no fixed timetable has been established, the preparation for our second registration is actively underway. We are deep in the strategy phase, considering various approaches to improve and adapt our subnet code. A critical decision looms — whether to further develop our existing human verification protocol or to shift towards an entirely new strategy for our next registration attempt. This period of brainstorming and planning is crucial as we aim to align more closely with the network’s expectations and our community’s needs.

Treasury Management

Registering on the Bittensor network involves locking a specific amount of TAO, which is not a payment but a commitment. For our registration as Subnet 38, a significant amount of TAO was locked. Following our deregistration, these funds were returned and subsequently transferred from the subnet owner’s wallet back to our original treasury wallet, thus consolidating our financial resources. The TAO has been redelegated with one of the top validators known for its high yields, ensuring our treasury continues to grow while we enhance the Taτsu Identity Protocol.

Maintaining Presence and Engagement

Despite the challenges on the mainnet, Taτsu’s involvement within the Bittensor community remains. We continue to maintain an active channel in the official Bittensor Discord, now positioned within the testnet section. Our continued involvement is supported by our active subnet on the testnet, Subnet 192. This engagement underscores our commitment to the Bittensor community and our ongoing role within the expansive TAO ecosystem, even as we prepare for our next steps on the mainnet.

As we navigate these changes and continue to evolve, we are grateful for the unwavering support from our community. The journey of Taτsu is filled with learning and adaptation, reflecting our dedication to growth and excellence in the decentralized AI landscape. We invite all members of the community to join us in this journey, participate in discussions, and stay updated on our progress through our active Discord presence. Together, we look forward to a future rich with innovation and success.

A Milestone Achievement

Last Friday marked a significant turning point for Taτsu as we achieved our most important milestone yet — launching a Bittensor subnet. The Taτsu Identity Protocol is now officially part of the Bittensor network, operating as Subnet 38. This achievement was the result of immense hard work, dedication, and unwavering commitment to our vision. Despite facing skepticism and numerous challenges, our team’s perseverance and integrity have led us to this success. You can explore our journey and current status on the TaoStats for Subnet 38.

The Beginning of a New Chapter

This launch does not signify an end, but rather the commencement of a new, promising era for Taτsu. The task now is to establish our worth within the TAO community — demonstrating our value, documenting our progress, and outlining our future plans. It is crucial now more than ever to gain the support of Bittensor miners and validators. Their engagement is essential for earning emissions, which will not only fund further development but also generate revenue to propel us forward.

Seizing Opportunities

The decision to register our subnet came sooner than anticipated. Originally, we had plans to release our updated whitepaper and engage extensively with validators and miners. However, when the cost of subnet creation dropped significantly — falling well below our budgeted allocation from the TAO treasury — we seized the opportunity. This proactive decision was made to capitalize on the favorable conditions and to mitigate potential cost increases due to the rising demand for subnet registrations.

We now invite all miners and validators to register on our subnet and run our code. Our GitHub hosts all the necessary technical guidelines and source code needed for participation.

Launching Whitepaper 2.0

In conjunction with these developments, we are excited to announce the release of our new Whitepaper 2.0. This document offers a comprehensive and detailed view of what Taτsu aims to achieve, including potential use cases and possible integrations. This whitepaper supersedes our previous Project Wiki. While the Wiki served primarily as a technical knowledge base, Whitepaper 2.0 delves deeper into our strategic vision, detailing our future plans, potential applications, and more. Those in search of technical guidelines will continue to find them updated on our GitHub.

You can access the full details of our renewed vision and roadmap in Tatsu Whitepaper 2.0.

Looking Ahead

As we forge ahead in this competitive space within the Bittensor network, our focus remains on innovation, community engagement, and strategic growth. The road ahead is filled with opportunities for expansion and deeper integration within the decentralized landscape.

We thank our community for their unwavering support and invite everyone to join us in this exciting new chapter of the Taτsu project. Stay tuned for more updates as we continue to evolve and strive for excellence in this new era of decentralization.