Thunder-LLM

Jaejin Lee's Thunder Research Group's Korean-English bilingual LLM: Llama 3.1 8B adapted via ~3TB of continual pretraining plus post-training, with the entire low-budget end-to-end pipeline — data collection, deduplication, the Thunder-Tok tokenizer, training, and evaluation — documented in the paper and released (weights CC-BY-NC-SA-4.0, toolkit under GitHub org mcrl). Averages 65.0 across nine Korean benchmarks vs 48.7 for the Llama 3.1 8B base while preserving English scores — a sovereign-AI-style recipe executed on an academic budget, evaluated on the group's own SNU_Ko-* suite. An instruction-tuned Thunder-LLM-Ins variant accompanies it.