Semantic Band-Gap (LLM + KG)
Fine-tuned two LoRA adapters on Llama 3.2 (3B) for a neuro-symbolic knowledge graph pipeline over semiconductor band-gap data (~150k materials, ~999k RDF triples, sourced from Materials Project, featurized via matminer). One adapter generates SPARQL from natural-language questions (eval loss 0.0004, 13/13 held-out queries parse-valid); the other extracts deduplicated triples from unstructured text and URLs (eval loss 0.0002, 21/21 valid JSON, 11/11 rejection cases correct). A deterministic sanitizer sits between the LLM and the graph — validates, repairs, and enforces schema constraints on every generated query, turning understood model failure modes into corrected output. RDF schema covers composition, crystal system, centrosymmetricity, band gap, and provenance. The graph serves as a single provenance-tracked source for band-gap screening, replacing manual cross-referencing of disconnected CSV, API, and literature sources. Query interface served as an API with interactive docs, available on request via email.
