Company Name (if applicable)

Project details

The MeTTa language is a multi-paradigm tool for declarative and functional computations over knowledge metagraphs, crucial in furthering the development of Artificial General Intelligence within the Hyperon framework. However, this powerful tool has an innovative nature and a unique syntax that makes it hard for new users to master. To address this challenge, this proposal focuses on creating a comprehensive MeTTa corpus – a high-quality dataset designed to train or fine-tune an AI model that will work as a coding assistant. Such coding assistants would be capable of generating correct, functional MeTTa code upon natural language input from a developer and simplify the way one learns and uses the language.

Beyond the corpus, this project will also provide open-source tools to automate the corpus creation process and extensive documentation to ensure transparency, reproducibility, and accessibility. Together, these efforts will lower barriers to MeTTa adoption, enhance usability, and accelerate AGI development within the Hyperon ecosystem. The resulting resources will not only serve immediate needs, but also give a sound foundation for further advances in the field.

Therefore, the project will be structured into three main milestones. The first milestone will be the creation of the MeTTa corpus itself, a high-quality dataset of 10,000 instruction-output pairs where instructions are expressed in natural language and their outputs are valid, error-free MeTTa code. This process will begin with data collection, leveraging resources such as the official MeTTa documentation, GitHub repositories, community-contributed code, and tutorials. To extract relevant data systematically, we will employ natural language processing (NLP) techniques, using libraries like SpaCy and NLTK, to analyze and parse instructional text, and web scraping frameworks like Scrapy for gathering examples from online sources. Data validation will be the next step, which will make sure that the dataset is free of errors, inconsistencies, and deprecated code. It will be performed by running syntax checking algorithms based on the grammar rules of MeTTa and testing frameworks that will verify the functional correctness of each code snippet. The checks will be implemented using technologies like PyTest and custom-written validators in Python.

In addition, when existing materials need supplementation to fill gaps, we will generate examples using generative approaches. Specifically, we will use fine-tuned language models to initially draft the instruction-output pairs, refining the output into MeTTa standards through some rule-based post-processing steps. In other words, combining data-driven generation and manual oversight will result in a highly diverse and representative dataset. Once the data is collected and validated, it will be categorized into sections such as basic constructs, advanced features, and practical applications. We will make use of technologies like Pandas for structuring data and JSON formats for storing the output to ensure usability. Finally, the entire corpus will undergo vigorous testing through automated tools that confirm syntactic and functional accuracy, making it a reliable resource for training AI models.

The second milestone entails the release of open-source tools and scripts that automate the processes involved in creating, validating, and extending the corpus. This will include data extraction scripts, which would systematically source code snippets and instructions from various resources, for which we will be using web scraping utilities like Selenium and Beautiful Soup, along with APIs where available. The validation tools will involve proprietary static code analysis algorithms to find syntactic errors and will use parsing and the validation of MeTTa-specific grammar by using frameworks such as Lark and ANTLR. Functional validation will involve testing MeTTa code snippets to verify their expected output in an isolated environment.

To supplement existing examples, users of the OSS will be able to use data synthesis tools leveraging transformer-based language models. Through prompt engineering, these tools will be able to generate instruction-output pairs that address advanced use cases and fill gaps in underrepresented features. These will be refined and validated with the help of rule-based algorithms. An evaluation framework using metrics will be developed, such as corpus completeness concerning key language feature coverage with accuracy rates. These will be implemented in Python and its libraries, NumPy and Matplotlib, used for quantitative analysis and reporting. All tools and scripts will be version-controlled through Git and hosted on GitHub to ensure openness and enable collaboration. This milestone ensures that the process of corpus creation is not only efficient, but also reproducible and scalable for future updates.

The third milestone will focus on extensive documentation, ensuring the project is transparent, easily replicable, and usable to the widest possible audience. This will describe the methodology employed in the corpus creation cycle, from data collection through validation, generation, and up to organization. This will also include instructions and example usages of open-sourced tools, designed to be accessible and easy to use for individuals of varying technical expertise. This documentation will also be detailed as to how this corpus addresses all MeTTa features, by way of examples, to make the breadth and depth of coverage clear. Any limitations will be acknowledged and recommendations for any future updates necessary to keep this corpus current with evolving language standards will be also discussed. Tutorials and practical guides on training or fine-tuning AI models using this corpus will also be provided in the documentation.

To sum up, this project will have three main deliverables, each designed to meet the immediate needs of MeTTa developers and to foster long-term AGI development within the Hyperon framework:

1.     Validated and Structured MeTTa Corpus:

A dataset of 10,000 instruction-output pairs, ensuring that MeTTa code generated is accurate and error-free. The corpus will span the complete range of MeTTa features, categorized into sections like basic constructs, advanced functionalities, and practical applications. Free and Open-Source Tools and Scripts:

2.     Automation of data extraction, validation, and corpus extension: The following will be provided:

- Data extraction scripts

- Validation: Static code analysis and functional testing

- Data synthesis: Synthesis of instruction-output pairs by transformer-based models

All these tools will be publicly available and thus can be replicated or extended.

3.     Comprehensive Documentation

Detailed documentation that covers all steps involved in the creation of the corpus, from data collection and validation to usage and generation. In particular, the following will be included:

- Step-by-step tutorials for new and experienced users alike

- Explanations of how the corpus addresses all MeTTa features

- Recommendations for future updates to maintain alignment with evolving standards

Moreover, to ensure sufficient material for the MeTTa corpus, we estimated the lines of code available across GitHub repositories, totalling 2,058,710 lines, excluding comments. After considering factors like relevance and usability, we project that 10-20% of this—approximately 205,871 to 411,742 lines—will be directly usable for natural language-to-MeTTa instruction-output pairs. With an average of 10 lines per pair, this would provide 20,587 to 41,174 pairs, well over the 10,000 needed for a high-quality dataset. This estimation confirms that the repositories provide a good basis for corpus creation while leaving room for refinement. The code for the estimation can be found in the references field.

This project represents a crucial step in lowering the barriers to the adoption of the MeTTa programming language. It will deliver a robust, validated corpus, automated tools, and comprehensive documentation that will create a scalable framework for empowering developers and researchers alike. In turn, these resources will accelerate the creation of an AI-powered coding assistant that simplifies the generation of correct and functional MeTTa code while broadening the language's accessibility. In the wider perspective, this project will contribute to the long-term development of Artificial General Intelligence with the enhanced ecosystem of Hyperon. The tools and methods developed in this area will provide a foundation for future innovation: they will enable sharing and collaboration to continuously improve. At its very core, the open-source principle of the project ensures that progress made today forms the basis for breakthroughs in AGI research tomorrow.

Our multidisciplinary team combines expertise in artificial intelligence, semantic data engineering, and software development, perfectly suited to create the MeTTa corpus and associated tools. Ilias Chrysovergis, Co-founder, CEO, and CTO of Metatopia, has led high-impact projects that bridge complex technical challenges with user-focused solutions. The last distinctive assignment which Ilias secured is a funding from the Greek Government (specifically from GRNET), to develop a specialized large language model oriented for Revit C# API. Anneza Bardani is a semantic data engineer, specializing in ontology design and data modeling. Dimitris Kleitsas specializes in machine learning and NLP. He has fine-tuned language models for automated content review. Iason Malkotsis experience in backend architecture will ensure efficient and reliable infrastructure for the tools and resources developed in the MeTTa ecosystem. Finally, Dimitra Pazouli has worked on LLM-based Level Generation for Games, showing her capabilities of using AI for creative and technical challenges.

Open Source Licensing

Links and references

Lines of Code analysis

https://metatopia.gr

https://metta-lang.dev/docs/learn

trueagi-io

vsbogd/metta-workshop-agi-24

logicmoo/metta-src-conversions

TownesZhou/hyperon-tutorials

Bonittas/MeTTa-language-Project

samwondim/Metta-programming

TollanBerhanu/metta-nl-query

Hope-Alemayehu/MeTTa_lang-Practices

Amanuel-1/metta-lang-highlighter

patham9/metta-nars

BlightG/meTTa_Training

logicmoo/vspace-metta

XBennyX2/Exercises-Metta

Nahom32/NMOSES

zarqa-ai/metta-motto

rejuve-bio/metta-motto-biochatter

Qoba-ai/MeTTa-KG