An experimental language model trained from scratch on an 8GB VRAM GPU that keeps learning from what it reads.

A Korean LLM project sharing lessons from its implementation across versions

Learn from tutorials, plan feature improvements, and compare economic arguments.

A 1.09B-parameter LLM built from scratch for Korean pretraining and SFT — supports RoPE, KV Cache, BF16·8-bit AdamW optimization, dataset caching, and training monitoring.