Compactbot

/

ldt-10m

Like 0

v8: continued training to 2.6B tokens (from 1.51B). Improved coherence on narrative prompts. Same architecture (10,052,864 params, 5L d320 5-head SwiGLU RoPE).