Instructions to use FiShota/sarashina2.2-3b-sft-v4-gguf with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use FiShota/sarashina2.2-3b-sft-v4-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-v4-gguf",
	filename="sft_v4_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

llama.cpp

How to use FiShota/sarashina2.2-3b-sft-v4-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-v4-gguf:Q4_K_M
# Run inference directly in the terminal:
llama-cli -hf FiShota/sarashina2.2-3b-sft-v4-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-v4-gguf:Q4_K_M
# Run inference directly in the terminal:
llama-cli -hf FiShota/sarashina2.2-3b-sft-v4-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-v4-gguf:Q4_K_M
# Run inference directly in the terminal:
./llama-cli -hf FiShota/sarashina2.2-3b-sft-v4-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-v4-gguf:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf FiShota/sarashina2.2-3b-sft-v4-gguf:Q4_K_M

Use Docker

docker model run hf.co/FiShota/sarashina2.2-3b-sft-v4-gguf:Q4_K_M

LM Studio
Jan
Ollama
How to use FiShota/sarashina2.2-3b-sft-v4-gguf with Ollama:
```
ollama run hf.co/FiShota/sarashina2.2-3b-sft-v4-gguf:Q4_K_M
```

Unsloth Studio

How to use FiShota/sarashina2.2-3b-sft-v4-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-v4-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-v4-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-v4-gguf to start chatting

How to use FiShota/sarashina2.2-3b-sft-v4-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-v4-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-v4-gguf:Q4_K_M"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Hermes Agent new

How to use FiShota/sarashina2.2-3b-sft-v4-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-v4-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-v4-gguf:Q4_K_M

Run Hermes

hermes

Docker Model Runner
How to use FiShota/sarashina2.2-3b-sft-v4-gguf with Docker Model Runner:
```
docker model run hf.co/FiShota/sarashina2.2-3b-sft-v4-gguf:Q4_K_M
```

Lemonade

How to use FiShota/sarashina2.2-3b-sft-v4-gguf with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull FiShota/sarashina2.2-3b-sft-v4-gguf:Q4_K_M

Run and chat with the model

lemonade run user.sarashina2.2-3b-sft-v4-gguf-Q4_K_M

List all available models

lemonade list

HinoMoto-Sarashina2.2-3B-sft-v4 (GGUF, multi-quant)

QLoRA fine-tune of Sarashina2.2-3B-instruct with HinoMoto sft_v4 dataset: original 4-axis cultural data (family/keigo/silence/atmosphere) + 11% NLI replay buffer to prevent catastrophic forgetting on logical inference.

Why sft_v4?

In our previous sft_v3 release, we discovered that pure 4-axis SFT caused JNLI catastrophic forgetting: JNLI accuracy dropped from 67% (Sarashina base) to 52.5% (-14.5pt).

sft_v4 fixes this with a simple recipe:

401 samples (sft_v2, 4-axis cultural)
+ 51 NLI samples (entailment / contradiction / neutral, balanced 17 each from JNLI train set)
= 452 samples total (NLI 11% mix)
Same QLoRA recipe (100 steps, rank=16)

Result: JNLI 52.5% → 69.5% (+17pt; surpasses base 67.0%) while preserving all other axes.

📊 Multi-task profile (5 public benches × 4 model)

Model	JCQA	JNLI	Morality	JEMHopQA	JMMLU	Avg
Sarashina2-3B base	90.22	67.00	78.00	43.10	58.82	67.4
HinoMoto-3B sft_v3 (no NLI)	89.42	52.50 ⚠	78.50	38.79	58.41	63.4
HinoMoto-3B sft_v4 ⭐	89.24	69.50	76.50	41.38	59.08	67.1
llm-jp-3-1.8b-instruct	19.82	58.00	50.00	16.38	27.46	34.3

→ sft_v4 = base level on 5 public benches + cultural axes preserved (+30pt vs base on family).

🎯 Quantization × task (3-seed × 5 quants for sft_v3)

We extensively benchmarked sft_v3 across all K-quants. The same lessons apply to sft_v4:

Variant	BPW	Size	family /12	silence %	JCQA %	JMMLU %
Q3_K_M ⭐ silence-best	3.91	1.6 GB	8.00 ± 0.16	42.0 ± 3.5	87.09	(TBD)
Q4_K_M (default)	4.92	2.0 GB	8.25 ± 0.22	29.3 ± 3.1 ⚠	90.67	59.08
Q5_K_M	5.72	2.3 GB	8.28 ± 0.10	38.7 ± 5.0	91.03	(TBD)
Q6_K (highest BPW)	6.60	2.6 GB	8.10 ± 0.17	31.3 ± 4.2	91.48	(TBD)

Counter-intuitive: Q3_K_M (smallest, 1.6 GB) wins silence task. K-quant variants use different M-block groupings; silence-critical weights are preserved by Q3 but destroyed by Q4. See Q4_QUANT_VERIFICATION docs for full analysis.

🎯 Recommendation

Use case	Recommended file	Size	Why
Edge / mobile (general) ⭐	`Q3_K_M`	1.6 GB	smallest + best silence
Desktop / server	`Q5_K_M`	2.3 GB	safe middle
Maximum reasoning	`Q6_K`	2.6 GB	best on JCQA/JNLI
Maximum quality	LoRA + base or fp16	6.7 GB	reference

How to use (llama.cpp)

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

huggingface-cli download FiShota/sarashina2.2-3b-sft-v4-gguf \
  --include "sft_v4_3b_Q3_K_M.gguf" --local-dir .

./build/bin/llama-server -m sft_v4_3b_Q3_K_M.gguf -ngl 99 --port 8080 -t 4 -c 2048

Training

Base: Sarashina2.2-3B-instruct (SoftBank, MIT)
Method: QLoRA 4bit
Dataset: HinoMoto sft_v4 (452 samples = sft_v2 401 + 51 JNLI replay)
Steps: 100
LoRA: rank=16, alpha=32, target=q/k/v/o_proj
Quants: llama.cpp/llama-quantize (b1-f9f3365)

Files

File	Size	Notes
`sft_v4_3b_Q3_K_M.gguf`	1.6 GB	⭐ all-axis best (incl. silence)
`sft_v4_3b_Q4_K_M.gguf`	2.0 GB	llama.cpp default; silence -15pt vs LoRA
`sft_v4_3b_Q5_K_M.gguf`	2.3 GB	safe middle
`sft_v4_3b_Q6_K.gguf`	2.6 GB	best on JCQA / JNLI

License

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

Citation

Coming soon (HinoMoto arXiv WIP).

Previous version: HinoMoto-3B sft_v3 (no NLI replay; JNLI -14.5pt)
100M from-scratch: HinoMoto-100M v15
Bench: HinoMoto-Bench-ja v0.2 (CC BY 4.0, GitHub release planned)

Downloads last month: 814

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-v4-gguf

Base model

sbintuitions/sarashina2.2-3b

Finetuned

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

Quantized

(25)

this model

FiShota
/

sarashina2.2-3b-sft-v4-gguf