Downloads · 30 days
66
10% of all-time downloads
FiShota/yamato-3b-v4-r64-legal-gguf
yamato-3b-v4-r64-legal-gguf is a machine learning model from FiShota. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as mit.
Best-performing Yamato variant on Yamato-Bench-ja v0.2 (n=30, nkw=92):
Downloads · 30 days
66
10% of all-time downloads
All-time downloads
650
Public
Repo size
15.6 GB
Likes
0
Public
Click a slice to open those files.
.gguf15.6 GB · 100%
From the Hugging Face model README
Best-performing Yamato variant on Yamato-Bench-ja v0.2 (n=30, n_kw=92):
| Variant | LoRA r | Overall | New domain | avg_len |
|---|---|---|---|---|
| v1 | 16 | 46.7% | 43.3% | 612 c |
| v2 | 16 (+10 samples) | 45.7% | 40.0% | 578 c |
| v3 | 32 | 48.9% | 43.3% | 501 c |
| v4 ⭐ | 64 | 54.3% | 53.3% | 415 c |
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:
| Variant | LoRA r | Bench % | Notes |
|---|---|---|---|
| v1 / v2 | 16 | 46.7% | baseline |
| v3 | 32 | 48.9% | +2.2 |
| v4 | 64 | 54.3% | +5.4 |
| v5 ⭐ | 128 | 57.6% | +3.3 (current best) |
→ All same recipe; only LoRA rank doubled. → Detailed analysis: HinoMoto 開発ノート #9 — "LoRA rank scaling boundary".