AI and ML research papers turned into working, reproducible implementations.
Every project takes a published paper and translates its core algorithms, experiments, and findings into clean, runnable code — with explanations that connect each implementation back to the paper.
Part of Syntran Labs
Paper Lab is a curated catalog of independent educational implementations of research papers from Syntran Labs. Each implementation lives in its own standalone paper-* repository with its own README, notebooks, tests, and CI.
Each project is built around one question:
"What would it look like if the paper's key ideas were implemented from scratch, clearly, and verified to work?"
That means every project in this collection aims to be:
| Quality | What It Means |
|---|---|
| 📖 Paper-grounded | Every class, function, and dataset maps back to a specific definition, algorithm, or section in the paper |
| 🔁 Reproducible | Key tables and figures from the paper are reproduced numerically and verified |
| 🪞 Independent | Not the official implementation — a clean, educational re-implementation with clear scope |
| ✅ Tested | Smoke-tests and CI confirm the implementation runs correctly |
| 🚫 No copyright violations | Paper PDFs are excluded; DOI references are provided in every repo |
Each project is a standalone repository with its own README, notebooks, tests, and GitHub Actions CI.
| Project | Paper | Year | Status | What It Reproduces |
|---|---|---|---|---|
| paper-rag-graph-4-datasets | Diamantini et al. — A Graph RAG Approach to Enhance Explainability in Dataset Discovery · DOI | 2026 | ✅ Published | Two-layer KG model (BKG + SKG), source profiling, KG-enriched query generation (Algorithm 1), preference-oriented ranking & explainability (Algorithm 2), end-to-end pipeline demo, Tables 1–5 · Python · Jupyter · NumPy · pandas · matplotlib · pytest |
Projects at an early stage — exploring methods, testbeds, and workflows that may eventually result in a published implementation.
| Project | Focus | Status |
|---|---|---|
| paper-eca-llm-hypothesis-workflow | Uses Elementary Cellular Automata as a bounded, reproducible testbed for governed LLM-assisted scientific hypothesis generation | 🌱 Early Incubation |
ECA LLM Hypothesis Workflow tests whether a governed LLM-assisted scientific workflow can generate falsifiable, reproducible, non-overclaiming hypotheses. The project uses the 256 Elementary Cellular Automata Wolfram rules as a finite and well-documented testbed, making overclaims easier to detect and follow-up experiments easier to reproduce. Governed by SYNTRAN AIEOS and methods-oriented by design — not a finished paper.
A paper implementation should make the reader understand the paper, not just run it.
For that reason, every project in this collection goes beyond code:
Working implementations Paper-grounded explanations Reproduced tables & figures
Offline mock outputs CI-validated smoke-tests Clear scope: what is and isn't reproduced
And follows a consistent structure:
paper-*/
├── README.md ← what paper, what is reproduced, how to run
├── CITATION.cff ← cite the implementation + the original paper
├── *.ipynb ← notebooks (one per major pipeline stage)
├── gen_notebooks.py ← source of truth; regenerates all notebooks
├── tests/ ← smoke-tests
└── .github/workflows/ ← CI
- Syntran Labs — The full portfolio
- learning-lab — Educational engineering projects track
- systems-lab — Production-oriented AI systems track
Built at Syntran Labs by Leonardo Sigales
Found an implementation useful? A ⭐ helps others find it too.