leonaej/WorldLeaf

Hand-curated African Savanna knowledge graph: 1,423 species, 1,541 ecological edges (predation, symbiosis, migration) sourced from iNaturalist, Wikidata & GloBI. Hosts a RAG baseline and an interpretable RL agent for multi-hop ecological reasoning. # WorldLeaf Hand-curated African Savanna knowledge graph with 1,423 species nodes and 1,541 ecological edges, built to support an interpretable reinforcement learning agent for multi-hop ecological reasoning. Master's thesis project. Author: Leona E Joseph Email: leonaej4000@gmail.com Affiliation: Rochester Institute of Technology --- ## Overview WorldLeaf constructs a structured knowledge graph of the Serengeti/African Savanna ecosystem and trains an RL agent to navigate it for multi-hop question answering. The project is motivated by the limitations of flat retrieval systems (RAG) on relational reasoning tasks, the RL agent is designed to exploit graph structure for interpretable, path-based answers. The project has three stages: - Stage 1: Dataset Construction - Stage 2: RAG Evaluation - Stage 3: RL Traversal Agent Training and Evaluation --- ## Knowledge Graph - 1,423 nodes: 760 iNaturalist species + 663 parent taxonomy nodes from Wikidata - 1,541 edges across 9 relation types: eats, preys_on, parasitizes, pollinates, parent_taxon, scavenges_from, migrates_with, disperses_seeds_of, symbiotic_with - Node attributes: node_id (Wikidata QID), name (scientific), common_name, rank, iconic_taxon - Stored as a NetworkX DiGraph (graph.gpickle) with CSV exports and an interactive Pyvis visualization --- ## Data Sources - iNaturalist: Species list for Serengeti National Park (place ID 69054) - Wikidata SPARQL:QID resolution, eats and parent_taxon edges - GloBI (Global Biotic Interactions): Preys_on, parasitizes, pollinates edges - Manual annotation: Scavenges_from, migrates_with, disperses_seeds_of, symbiotic_with edges - Wikipedia: Full article text per node for RAG baseline --- ## Project Structure ``` WORLDLEAF/ ├── src/dataset/ │ ├── wikidata_fetch.py # iNaturalist species collection and Wikidata edge fetching │ ├── globi_fetch.py # GloBI interaction edge fetching │ ├─ …