I trained a 113M-parameter earthquake LLM from absolute scratch

A developer trained a 113-million-parameter language model specialized in earthquake science from scratch, using six free data sources mixed with general text (24% domain-specific, 76% general) and documented the entire process from data crawling to inference. The model achieved 0.997 bits/byte fluency on general text—35% better than a domain-only baseline—and was trained on two NVIDIA A30 GPUs in approximately 6.5 hours, with the project designed as a teaching resource to explain each stage of language model pretraining. The nanoGPT-Seis repository demonstrates a complete LLM lifecycle workflow that can be replicated on consumer hardware.