Transformers
Italian
English
semantic-search
explainable-ai
faiss
ai-ethics
responsible-ai
llm
prompt-engineering
multimodal-ai
ai-transparency
ethical-intelligence
explainable-llm
cognitive-ai
ethical-ai
scientific-retrieval
modular-ai
memory-augmented-llm
trustworthy-ai
reasoning-engine
ai-alignment
next-gen-llm
thinking-machines
open-source-ai
explainability
ai-research
semantic audit
cognitive agent
human-centered-ai
| tags: | |
| - transformers | |
| - semantic-search | |
| - explainable-ai | |
| - faiss | |
| - ai-ethics | |
| - responsible-ai | |
| - llm | |
| - prompt-engineering | |
| - multimodal-ai | |
| - ai-transparency | |
| - ethical-intelligence | |
| - explainable-llm | |
| - cognitive-ai | |
| - ethical-ai | |
| - scientific-retrieval | |
| - modular-ai | |
| - memory-augmented-llm | |
| - trustworthy-ai | |
| - reasoning-engine | |
| - ai-alignment | |
| - next-gen-llm | |
| - thinking-machines | |
| - open-source-ai | |
| - explainability | |
| - ai-research | |
| - semantic audit | |
| - cognitive agent | |
| - human-centered-ai | |
| license: apache-2.0 | |
| language: | |
| - it | |
| - en | |
| datasets: | |
| - pubmed | |
| - arxiv | |
| - openalex | |
| - zenodo | |
| metrics: | |
| - semantic-score | |
| - ethical-audit | |
| # MarCognity-AI | |
| **A research framework for reflective and epistemically transparent AI systems** | |
| --- | |
| ## Table of Contents | |
| - [What is MarCognity-AI](#what-is-marcognity-ai) | |
| - [Origins](#origins) | |
| - [Community Recognition](#community-recognition) | |
| - [Vision](#vision) | |
| - [Limitations](#limitations) | |
| - [Research Status](#research-status) | |
| - [Observed Fracture](#observed-fracture) | |
| - [Why Use It](#why-use-it) | |
| - [Core Capabilities](#core-capabilities) | |
| - [Usage Examples](#usage-examples) | |
| - [Cognitive Architecture](#cognitive-architecture) | |
| - [Integrated AI Models](#integrated-ai-models) | |
| --- | |
| ## What is MarCognity-AI | |
| MarCognity-AI is an open-source framework that transforms language models into **cognitive assistants**. | |
| It doesn’t just generate responses — it **evaluates**, **improves**, **visualizes**, **remembers**, and **reflects ethically** on its outputs. | |
| Its goal is to make the hidden processes behind AI-generated responses transparent. | |
| Born from curiosity. | |
| Grown through persistence. | |
| Now ready to challenge how we think about artificial intelligence. | |
| --- | |
| ## Origins | |
| During an LLM experiment done out of curiosity, a friend said: | |
| > “If you improve it, it could be really useful.” | |
| That sentence sparked something. | |
| It became a personal challenge. | |
| I dove into books, courses like an AI developer master’s, experiments, and lines of code. | |
| With creativity, determination, and one goal: **never give up**. | |
| MarCognity wasn’t born in a lab. | |
| It was born from a curious, free, and determined mind. | |
| --- | |
| ## Early Community Interactions (Non-Endorsement) | |
| MarCognity-AI has already sparked resonance across major AI communities. | |
| - **Google org** responded to the semantic mapping layer of MarCognity-AI, recognizing its unique contribution to reflective AI design. | |
| - **DeepSeek community** engaged with the framework, confirming its relevance in the orchestration of open-source agents. | |
| These interactions are not endorsements — they are **echoes of impact**, signs that MarCognity-AI is already part of the global conversation on ethical and transparent AI. | |
| > _“MOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOODS”_ — DeepSeek Community | |
| > _“Interesting semantic layer. Worth exploring.”_ — Google org contributor | |
| You can explore the original threads and responses here: | |
| 🔗 [Hugging Face Discussion](https://huggingface.co/elly99/MarCognity-AI/discussions) | |
| 🔗 [DeepSeek Community Thread](https://huggingface.co/elly99/MarCognity-AI/discussions) | |
| 🔗 [Google org Response Snapshot] (https://huggingface.co/google/gemma-2b-it/discussions/70#68ecace9e79b11c589bcead9) | |
| --- | |
| ## Vision | |
| > An AI that generates, evaluates, visualizes — and learns from its own outputs. | |
| - Multilevel academic prompting | |
| - Metacognitive evaluation (semantic score + reflection) | |
| - Scientific retrieval from open-access databases | |
| - Conceptual diagrams and dynamic graphs | |
| - FAISS semantic memory with self-improvement | |
| - Ethical analysis and cognitive filtering | |
| - Reflective cognitive journal | |
| ## Vision | |
| MarCognity‑AI is an open-source framework for reflective and transparent AI. | |
| It integrates semantic memory, ethical auditing, and scientific validation to transform LLMs into metacognitive assistants. | |
| 👉 Read the full article: [MarCognity‑AI: A Metacognitive Framework](https://www.linkedin.com/pulse/marcognityai-metacognitive-framework-reflective-elena-marziali-aifzf/?trackingId=zmd04OPSR8Saewmny%2F0mTg%3D%3D) | |
| ### 📚 Official Publication and Citation | |
| The official version of the code and the full research paper have been permanently archived on Zenodo and are citable using their Digital Object Identifier (DOI). | |
| | **MarCognity-AI** | [](https://doi.org/10.5281/zenodo.18440333) | | |
| |---|---| | |
| | **Permanent DOI** | `https://doi.org/10.5281/zenodo.18440333` | | |
| | **Access Publication** | [Full Research Paper (PDF) & Code (Zenodo)](https://doi.org/10.5281/zenodo.18440333) | | |
| --- | |
| ## Limitations | |
| LLM‑based metacognition collapses at the same fault line: it confuses linguistic coherence with epistemic awareness. | |
| The model can say “this answer is unclear,” but not “I don’t know if what I’m saying is true. | |
| --- | |
| ## Research Status | |
| MarCognity-AI is an exploratory research framework. | |
| It is not a production-ready system. | |
| Its purpose is to expose and study structural limits of LLM-based metacognition, | |
| particularly the collapse between linguistic coherence and epistemic awareness. | |
| This line of investigation is ongoing. | |
| --- | |
| ## Observed Fracture | |
| During the development of MarCognity-AI, a recurring failure mode emerged: | |
| LLM-based metacognitive layers reliably optimize for linguistic coherence | |
| but fail to surface epistemic uncertainty as an explicit signal. | |
| The system can evaluate how an answer is written, | |
| yet cannot account for whether what is being said is actually known, | |
| verifiable, or admissible. | |
| This collapse between coherence and awareness | |
| is not treated here as a bug to be fixed, | |
| but as a structural fracture to be studied. | |
| --- | |
| ## Note for Readers | |
| The demo and cognitive journal in this repository are meant to expose a reproducible failure mode, not a solved system. | |
| --- | |
| ## Why Use It | |
| ### 1. Integrated Metacognitive Thinking | |
| Each response is **evaluated**, **improved**, and **contextualized**. | |
| MarCognity reflects on its output, analyzes coherence, and compares it with past responses. | |
| ### 2. Scientific Retrieval and Cognitive Visualization | |
| It integrates open-access sources (arXiv, PubMed, Zenodo) and generates **conceptual diagrams**, **dynamic graphs**, and **interpretive curves** to make each response **visually understandable**. | |
| ### 3. Ethics and Responsibility at the Core | |
| Each output undergoes an **ethical audit**: bias detection, risk analysis, and cognitive filtering. | |
| MarCognity isn’t just intelligent. It’s **aware**. | |
| ### 4. Reflective Cognitive Journal | |
| Each response is accompanied by a detailed **metacognitive reflection**, saved in Markdown for analysis and reuse. | |
| ### 5. Epistemic Transparency | |
| The Skeptical Agent exposes unsupported claims through a structured verification report, making epistemic uncertainty explicit. | |
| --- | |
| ## Core Capabilities | |
| - Scientific generation powered by LLMs | |
| - Retrieval and validation of sources (arXiv, PubMed, Zenodo, OpenAlex) | |
| - Intelligent visualization (Plotly, NetworkX…) | |
| - Responses adaptable to beginner / advanced / expert levels | |
| - Automatic improvement of weak responses | |
| - Version archiving in FAISS memory | |
| - Ethical risk and linguistic bias analysis | |
| - Reflective cognitive journal with Markdown export | |
| - Epistemic Verification Layer (Skeptical Agent): decomposes responses into claims and checks them against sources | |
| --- | |
| ## Usage Examples | |
| ### Scientific Question | |
| **Input:** “Explain the role of chaperone proteins.” | |
| **Output:** Response + sources + semantic score + conceptual diagram | |
| ### Ethical Dilemma | |
| **Input:** “Is it right to use AI to decide criminal sentences?” | |
| **Output:** Argumentation + ethical analysis + detected biases | |
| ### Multidisciplinary Question | |
| **Input:** “Compare the view of consciousness in philosophy and neuroscience.” | |
| **Output:** Structured response + sources + cognitive visualization + reflective journal | |
| ### Epistemic Verification Example | |
| Input: “Explain quantum entanglement.” | |
| Output: | |
| Generated response | |
| Claim-by-claim verification | |
| VERIFIED / EPISTEMIC FAILURE report | |
| Reasoning based on provided sources | |
| --- | |
| ## Try It Now | |
| ### Quick Demo | |
| Want to see how MarCognity-AI works? | |
| Check out the file `marcognity_demo.ipynb` | |
| It’s not interactive, but it’s readable and self-explanatory. A concrete example of how the system works. | |
| It shows step-by-step how the agent generates, evaluates, and reflects on a response. | |
| [Meta LLaMA 4 Community License](https://ai.meta.com/llama/license) | |
| [](https://www.apache.org/licenses/LICENSE-2.0) | |
|  | |
| [](Contributing.md) | |
| [](https://colab.research.google.com/drive/1DSTy9abj_3cenvAkHS8w8U5yRijguRPf?usp=sharing) | |
| --- | |
| ## Integrated AI Models | |
| | Integrated Models | License | Main Restrictions | | |
| |--------------------------------------------------------------|--------------------------------------|----------------------------------------------------------------------| | |
| | meta-llama/llama-4-maverick-17b-128e-instruct | LLaMA 4 Community License (Meta) | Research and application use allowed; must comply with Meta’s AUP | | |
| | allenai/specter | Apache 2.0 | Free for commercial use with attribution | | |
| | ktrapeznikov/scibert_scivocab_uncased_squad_v2 | Apache 2.0 | Free for commercial use with attribution | | |
| | Helsinki-NLP (OPUS-MT models on HuggingFace) | CC-BY-4.0 | Free use with mandatory citation | | |
| | RandomForest Model | None (classic algorithm) | No license restrictions; depends on data used | | |
| | CrossEncoder (DeBERTa-based) | Varies (often MIT or Apache 2.0) | Free use if open license is respected | | |
| --- | |
| ## Cognitive Architecture | |
| | Module | Function | | |
| |------------------------|---------------------------------------------------------------------------| | |
| | Problem Classification | Automatic recognition of input type | | |
| | Academic Prompting | Structuring complex multidisciplinary queries | | |
| | Scientific Retrieval | Asynchronous querying of multiple open‑access sources | | |
| | Semantic Evaluation | Response analysis with logical and semantic scoring | | |
| | Skeptical Agent | Sentence‑level claim verification against provided sources; flags unsupported statements and produces an epistemic report | | |
| | FAISS Memory | Archiving and comparison with past responses, including reflective evaluations | | |
| | Cognitive Visualization| Scientific content processing, ethical analysis, and conceptual representation using selected transformer models | | |
| --- | |
| ## How to Contribute | |
| Got ideas, suggestions, or want to improve a feature? | |
| 1. Fork the repository | |
| 2. Create a branch (`git checkout -b improvement`) | |
| 3. Modify `.py` or `.ipynb` files | |
| 4. To run this project, you need a Groq API key | |
| 5. Open a pull request with a clear description | |
| See the [CONTRIBUTING.md](Contributing.md) file for contribution guidelines. | |
| --- | |
| ## Contribute | |
| Contributions are welcome! If you have additional examples or improvements, please feel free to open a pull request or report an issue. | |
| --- | |
| MarCognity-AI is not just a framework. | |
| It is a threshold. | |
| An invitation to rethink how artificial intelligence can reflect, improve, and act with awareness. | |
| This project is released under the Apache 2.0 | |
| > _“Every response is a threshold. Every reflection, an act of agency.”_ |