Best-performing Yamato variant on Yamato-Bench-ja v0.2 (n=30, n_kw=92):

Same 60 hand-crafted legal Q&A samples. Only LoRA rank changed. Rank scaling at small-domain SFT: r=16 → r=32 → r=64 yields monotonic gains (+2.2pt, +5.4pt). avg_len contracts as rank grows = tighter on-task focus.

Use this if you want maximum quality. v3 (r=32) for balanced size/speed, v1/v2 for sample-efficient baseline.

huggingface-cli download FiShota/yamato-3b-v4-r64-legal-gguf yamato_3b_v4_r64.Q4_K_M.gguf --local-dir .
~/projects/llama.cpp/build/bin/llama-server -m yamato_3b_v4_r64.Q4_K_M.gguf -c 2048 -ngl 99 --port 8092

MIT (inherits from Sarashina2.2-3B-instruct-v0.1).

This is one variant in a 4-point LoRA rank ablation on the same 60 legal samples:

→ All same recipe; only LoRA rank doubled. → Detailed analysis: HinoMoto 開発ノート #9 — "LoRA rank scaling boundary".