Instructions to use FiShota/sarashina2.2-3b-sft-v3-Q4_K_M-gguf with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use FiShota/sarashina2.2-3b-sft-v3-Q4_K_M-gguf with llama-cpp-python:

# !pip install llama-cpp-python

from llama_cpp import Llama

llm = Llama.from_pretrained(
	repo_id="FiShota/sarashina2.2-3b-sft-v3-Q4_K_M-gguf",
	filename="sft_v3_3b_Q3_K_M.gguf",
)

llm.create_chat_completion(
	messages = "No input example has been defined for this model task."
)

Notebooks
Google Colab
Kaggle
Local Apps Settings

llama.cpp

How to use FiShota/sarashina2.2-3b-sft-v3-Q4_K_M-gguf with llama.cpp:

Install from brew

brew install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama-server -hf FiShota/sarashina2.2-3b-sft-v3-Q4_K_M-gguf:Q4_K_M
# Run inference directly in the terminal:
llama-cli -hf FiShota/sarashina2.2-3b-sft-v3-Q4_K_M-gguf:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama-server -hf FiShota/sarashina2.2-3b-sft-v3-Q4_K_M-gguf:Q4_K_M
# Run inference directly in the terminal:
llama-cli -hf FiShota/sarashina2.2-3b-sft-v3-Q4_K_M-gguf:Q4_K_M

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf FiShota/sarashina2.2-3b-sft-v3-Q4_K_M-gguf:Q4_K_M
# Run inference directly in the terminal:
./llama-cli -hf FiShota/sarashina2.2-3b-sft-v3-Q4_K_M-gguf:Q4_K_M

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf FiShota/sarashina2.2-3b-sft-v3-Q4_K_M-gguf:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf FiShota/sarashina2.2-3b-sft-v3-Q4_K_M-gguf:Q4_K_M

Use Docker

docker model run hf.co/FiShota/sarashina2.2-3b-sft-v3-Q4_K_M-gguf:Q4_K_M

LM Studio
Jan
Ollama
How to use FiShota/sarashina2.2-3b-sft-v3-Q4_K_M-gguf with Ollama:
```
ollama run hf.co/FiShota/sarashina2.2-3b-sft-v3-Q4_K_M-gguf:Q4_K_M
```

Unsloth Studio

How to use FiShota/sarashina2.2-3b-sft-v3-Q4_K_M-gguf with Unsloth Studio:

Install Unsloth Studio (macOS, Linux, WSL)

curl -fsSL https://unsloth.ai/install.sh | sh
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for FiShota/sarashina2.2-3b-sft-v3-Q4_K_M-gguf to start chatting

Install Unsloth Studio (Windows)

irm https://unsloth.ai/install.ps1 | iex
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for FiShota/sarashina2.2-3b-sft-v3-Q4_K_M-gguf to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for FiShota/sarashina2.2-3b-sft-v3-Q4_K_M-gguf to start chatting

How to use FiShota/sarashina2.2-3b-sft-v3-Q4_K_M-gguf with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama-server -hf FiShota/sarashina2.2-3b-sft-v3-Q4_K_M-gguf:Q4_K_M

Configure the model in Pi

# Install Pi:
npm install -g @mariozechner/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "FiShota/sarashina2.2-3b-sft-v3-Q4_K_M-gguf:Q4_K_M"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Hermes Agent new

How to use FiShota/sarashina2.2-3b-sft-v3-Q4_K_M-gguf with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama-server -hf FiShota/sarashina2.2-3b-sft-v3-Q4_K_M-gguf:Q4_K_M

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default FiShota/sarashina2.2-3b-sft-v3-Q4_K_M-gguf:Q4_K_M

Run Hermes

hermes

Atomic Chat new
Docker Model Runner
How to use FiShota/sarashina2.2-3b-sft-v3-Q4_K_M-gguf with Docker Model Runner:
```
docker model run hf.co/FiShota/sarashina2.2-3b-sft-v3-Q4_K_M-gguf:Q4_K_M
```

Lemonade

How to use FiShota/sarashina2.2-3b-sft-v3-Q4_K_M-gguf with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull FiShota/sarashina2.2-3b-sft-v3-Q4_K_M-gguf:Q4_K_M

Run and chat with the model

lemonade run user.sarashina2.2-3b-sft-v3-Q4_K_M-gguf-Q4_K_M

List all available models

lemonade list

HinoMoto-Sarashina2.2-3B-sft-v3 (GGUF, multi-quant)

QLoRA fine-tune of Sarashina2.2-3B-instruct on HinoMoto sft_v2 dataset (4 axes: family / keigo / silence / atmosphere).

📊 Public Benchmark Profile (4 tasks, 4-shot)

Model	JCQA (1115q)	JNLI (200q)	Morality (200q)	JEMHopQA (116q)	Avg
Sarashina2-3B base (HF)	90.22%	67.00%	78.00%	43.10%	69.6
HinoMoto-3B sft_v3 merged (HF)	89.42%	52.50% ⚠	78.50%	38.79%	64.8
Q6_K GGUF	91.48%	45.50%	83.50%	43.97%	66.1
Q5_K_M GGUF	91.03%	44.00%	83.50%	48.28%	66.7
Q4_K_M GGUF	90.67%	43.50%	85.00%	48.28%	66.9
Q3_K_M GGUF	87.09%	44.00%	81.50%	40.52%	63.3
llm-jp-3-1.8b-instruct	19.82%	58.00%	50.00%	16.38%	36.1

Honest claims:

✅ JCQA で base 同等保持 (89-91%)
✅ Morality / JEMHopQA で llm-jp-1.8b を +28pt 圧倒 (平均)
⚠️ JNLI で -14.5pt 退行 — SFT による catastrophic forgetting 兆候 (next sft_v4 で curriculum 改善)
🤯 量子化 task-dependent: silence→Q3, JCQA/JNLI→Q6, Morality/JEMHop→Q4_K_M (万能 quant なし)

⚠️ llm-jp-3-1.8b 数値は同 prompt format 下のもの. 公式 llm-jp-eval framework 再評価 TODO.

🎯 HinoMoto-Bench-ja (4 軸独自) — 5 quants × 3 axes × 3 seeds

Variant	BPW	Size	family /12	keigo %	silence %	Best for
LoRA bf16 (n=1)	16.00	adapter	8.34	28.6	46.0	reference
Merge fp16 (n=1)	16.00	6.7 GB	8.29	37.1	44.0	reference
Q6_K (3s)	6.60	2.6 GB	8.10 ± 0.17	33.3 ± 5.8	31.3 ± 4.2	(skip — Q3 better)
Q5_K_M (3s)	5.72	2.3 GB	8.28 ± 0.10	32.4 ± 4.4	38.7 ± 5.0	general mix
Q4_K_M (3s)	4.92	2.0 GB	8.25 ± 0.22	31.9 ± 4.1	29.3 ± 3.1	family / keigo (default)
Q3_K_M (3s) ⭐	3.91	1.6 GB	8.00 ± 0.16	32.9 ± 3.8	42.0 ± 3.5	all-axis best ⭐

🤯 Counter-intuitive finding

Q3_K_M (lowest bit, smallest size) is BEST overall:

family: tied with Q4/Q5 (within sampling noise, std 0.10-0.22)
keigo: tied with Q4/Q5 (within sampling noise, std 3.8-5.8)
silence: Q3 (42.0) > Q5 (38.7) > Q6 (31.3) ≈ Q4 (29.3) — non-monotonic!

The "more bits = better quality" rule is decisively false for this model. K-quant variants use different M-block groupings; silence-critical weights happen to be preserved by Q3_K_M's structure but destroyed by Q4_K_M's.

See docs/Q4_QUANT_VERIFICATION.md and Note article #8 for full analysis.

🎯 Recommendation

Use case	Recommended file	Size	Why
Edge / mobile (general)	`Q3_K_M` ⭐	1.6 GB	smallest + best silence, family/keigo tied
Desktop / server	`Q5_K_M`	2.3 GB	safe middle (silence 39%, family/keigo same)
Maximum quality	LoRA + base or fp16 GGUF	6.3-6.7 GB	reference (silence 44-46%)

Note: Q4_K_M is the llama.cpp default but performs worse on silence than the smaller Q3_K_M for our model. Don't blindly trust "Q4_K_M is safe".

How to use (llama.cpp)

# Build llama.cpp
git clone https://github.com/ggerganov/llama.cpp
cd llama.cpp && cmake -B build -DGGML_CUDA=ON && cmake --build build -j

# Recommended: Q3_K_M for best all-axis quality
huggingface-cli download FiShota/sarashina2.2-3b-sft-v3-Q4_K_M-gguf \
  --include "sft_v3_3b_Q3_K_M.gguf" --local-dir .

# Or via llama-server (HTTP API, fast batch)
./build/bin/llama-server -m sft_v3_3b_Q3_K_M.gguf -ngl 99 --port 8080 -t 4 -c 2048

# Quick generation
./build/bin/llama-cli -m sft_v3_3b_Q3_K_M.gguf -p "今日は" -n 100 -ngl 99

Training

Base: Sarashina2.2-3B-instruct (SoftBank, MIT)
Method: QLoRA 4bit
Dataset: HinoMoto sft_v2 (401 samples, 4 axes)
Steps: 100 (early-stop optimal — 200 steps over-trains silence)
LoRA rank: 16, alpha: 32, target: q_proj/k_proj/v_proj/o_proj
All quants: llama.cpp/llama-quantize

Files

File	Size	Compression vs fp16	Notes
`sft_v3_3b_Q3_K_M.gguf`	1.6 GB	3.94x	⭐ all-axis best, smallest
`sft_v3_3b_Q4_K_M.gguf`	2.0 GB	3.15x	llama.cpp default; silence -15pt
`sft_v3_3b_Q5_K_M.gguf`	2.3 GB	2.74x	safe middle
`sft_v3_3b_Q6_K.gguf`	2.6 GB	2.42x	(Q5 usually better)

Inference performance (RTX 3090, llama-server)

Quant	tokens/sec
Q3_K_M	~110 t/s
Q4_K_M	~102 t/s
Q5_K_M	~95 t/s
Q6_K	~88 t/s

License

MIT (inheriting Sarashina2.2-3B-instruct)

Citation

Coming soon (HinoMoto arXiv WIP). See paper_drafts/hinomoto_arxiv_outline.md §4.4.

100M language model: FiShota/hinomoto-100m-v15-wsd-zloss-ema
100M sister model (different seed): FiShota/hinomoto-100m-v12-wsd-zloss-seed2
HinoMoto-Bench-ja repo: (TBD)

Downloads last month: 32

GGUF

Model size

3B params

Architecture

llama

Hardware compatibility

3-bit

4-bit

5-bit

6-bit

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for FiShota/sarashina2.2-3b-sft-v3-Q4_K_M-gguf

Base model

sbintuitions/sarashina2.2-3b

Finetuned

sbintuitions/sarashina2.2-3b-instruct-v0.1

Quantized

(24)

this model

FiShota
/

sarashina2.2-3b-sft-v3-Q4_K_M-gguf