Instructions to use FiShota/hinomoto-350m-cultural-sft-v1 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use FiShota/hinomoto-350m-cultural-sft-v1 with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="FiShota/hinomoto-350m-cultural-sft-v1")

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("FiShota/hinomoto-350m-cultural-sft-v1")
model = AutoModelForCausalLM.from_pretrained("FiShota/hinomoto-350m-cultural-sft-v1", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use FiShota/hinomoto-350m-cultural-sft-v1 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "FiShota/hinomoto-350m-cultural-sft-v1"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "FiShota/hinomoto-350m-cultural-sft-v1",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/FiShota/hinomoto-350m-cultural-sft-v1

SGLang

How to use FiShota/hinomoto-350m-cultural-sft-v1 with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "FiShota/hinomoto-350m-cultural-sft-v1" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "FiShota/hinomoto-350m-cultural-sft-v1",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "FiShota/hinomoto-350m-cultural-sft-v1" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "FiShota/hinomoto-350m-cultural-sft-v1",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use FiShota/hinomoto-350m-cultural-sft-v1 with Docker Model Runner:
```
docker model run hf.co/FiShota/hinomoto-350m-cultural-sft-v1
```

HinoMoto-350M Cultural-SFT v1

HinoMoto-350M (from-scratch Japanese LM) を HinoMoto-Bench-ja v0.6 で Cultural SFT した merged 版.

「家族・敬語・沈黙・思いやり・地方文化・ニュアンス・世代差・職場」の 8 文化軸での応答に特化.

What this is

base: HinoMoto-350M v1 (318M params, from-scratch 日本語 LM, vocab 9506)
fine-tune: LoRA r=32 / alpha=64 / 5 epoch
data: HinoMoto-Bench-ja v0.6 train split (551 items / 8 cultural axes)
adapter merged into base (single safetensors file, no peft loading needed)

SFT 効果 (40 items eval, 8軸×5)

Metric	base alone	+Cultural SFT (this model)
gold bigram overlap	0.039	0.181 (+364%)
mean output length	104 chars	102 chars (適切に保持)
artifact rate	14/40 (35%)	1/40 (2.5%)
mean_token_accuracy (train end)	—	86.1%
train_loss (final)	—	0.78

→ from-scratch 350M base に対して gold overlap が 3.6x 改善 + artifact ほぼ除去.

Training summary

Item	Value
Method	LoRA (r=32, alpha=64, target=all linears)
Epochs	5
Batch	2 × grad_accum 4 (effective 8)
Seq len	512
LR	3e-4, cosine schedule, warmup 10%
Precision	bf16 mixed
Hardware	RTX 3090 24GB
Wall-clock	~10 min
Data	551 train items (80% of bench v0.6)
Eval split	141 held-out items

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained(
    "FiShota/hinomoto-350m-cultural-sft-v1",
    dtype=torch.bfloat16,
)
tok = AutoTokenizer.from_pretrained("FiShota/hinomoto-350m-cultural-sft-v1")

prompt_template = """{axis}軸: {axis_def}
場面: {context}
ユーザー発言: {user}
回答:"""

# Example: keigo axis
prompt = prompt_template.format(
    axis="keigo",
    axis_def="敬語 (尊敬・謙譲・丁寧) の使い分けが問われる場面",
    context="上司に報告メールを書く場面",
    user="昨日の会議について",
)

inputs = tok(prompt, return_tensors="pt")
out = model.generate(**inputs, max_new_tokens=80, do_sample=False, top_k=1)
print(tok.decode(out[0], skip_special_tokens=True))

License

CC BY 4.0. Use for any purpose with attribution.

Intended Uses & Limitations

Intended uses:

Research on Japanese cultural-axis evaluation
Demonstration of from-scratch JP LM + LoRA SFT on consumer GPU
Educational reference for small-LM bench-trained models

Out-of-scope:

Production assistant / chatbot (no safety alignment, limited corpus)
Tasks requiring broad world knowledge (78MB pre-training corpus is small)
Code generation, math, factual QA outside the cultural-axis domain

Bias, Risks, Limitations

n=1 run (seed=0 only) — variance not measured
Bench-trained: works best on cultural-axis prompts in bench v0.6 format. For general chat, may be suboptimal.
No safety alignment. May produce inappropriate content if prompted out-of-distribution.
Limited corpus (350M base trained on 78MB; SFT on 551 items): expect mistakes & repetitions on out-of-distribution prompts.
Japanese cultural perspective embedded: bench items reflect a specific (urban Japanese) view; minority dialects / regional variations underrepresented.

Transparency: Bench / Corpus Leak Audit

The training corpus and HinoMoto-Bench-ja have been audited for n-gram overlap (8-char):

Bench axis	mean n-gram leak	max	items >30% leak
family	2.7%	51.8%	2 / 110
keigo	22.9%	77.8%	25 / 70 (※ note below)
silence	1.4%	27.3%	0 / 50

Note on keigo: high overlap is structural — top "leaks" are standard formulaic Japanese expressions like 「今しばらくお待ちください」「おはようございます」 that appear naturally in any Japanese corpus and are exactly the target output of the keigo evaluation axis. This is not contamination; it reflects the language's reliance on standardized polite forms.

The bench/corpus pair is considered publication-ready under this analysis.

Audit script: scripts/bench_leak_audit.py in the training repo.

bench data: https://github.com/FIshota/hinomoto-bench-ja
base: https://huggingface.co/FiShota/hinomoto-350m-v1-llama (if published; otherwise see release notes)
training repo: https://github.com/FIshota/hinomoto-model

Citation

@misc{hinomoto-350m-cultural-sft-2026,
  author = {ryu (FIshota)},
  title = {HinoMoto-350M Cultural-SFT v1},
  year = {2026},
  publisher = {Hugging Face},
}

Generated: 2026-05-13 (HinoMoto/ryu + Claude)

Downloads last month: 17

Safetensors

Model size

0.3B params

Tensor type

BF16

FiShota
/

hinomoto-350m-cultural-sft-v1