from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch

base = "sbintuitions/sarashina2.2-3b-instruct-v0.1"
adapter = "FiShota/hinomoto-family-3b-v3"

tok = AutoTokenizer.from_pretrained(adapter)
model = AutoModelForCausalLM.from_pretrained(
    base, dtype=torch.bfloat16, device_map="cuda", trust_remote_code=True,
)
model = PeftModel.from_pretrained(model, adapter)
model.eval()

messages = [
    {"role": "system", "content": "あなたは家族向けの日本語 AI アシスタント「ai-chan」です。"},
    {"role": "user", "content": "## 文脈\n平日の朝、子供を起こす場面\n\n## ユーザー発言\n(まだ眠そう) おはよう..."},
]
inputs = tok.apply_chat_template(messages, return_tensors="pt", add_generation_prompt=True, tokenize=True).to("cuda")
out = model.generate(inputs, max_new_tokens=80, temperature=0.7, top_p=0.9, do_sample=True,
                     pad_token_id=tok.eos_token_id)
print(tok.decode(out[0][inputs.size(1):], skip_special_tokens=True))

Sample output: 「おはよう！今日も一日頑張ろうね。」

Training

QLoRA training:

bash

python scripts/sft_qlora_messages.py \
  --model sbintuitions/sarashina2.2-3b-instruct-v0.1 \
  --data data/sft_v2.jsonl \
  --out artifacts/sft_v3_3b_qlora \
  --max-steps 100 --batch-size 1 --grad-accum 8

Limitations

401 SFT samples のみ (production には ~10k+ 推奨)
単一 seed (= 0) のみ、variance 未測定
HinoMotoBench-ja は own bench, third-party annotator 未実施
7B+ base への extension は未実装 (将来 KAGUYA-9B で実装予定)

Citation

bibtex

@misc{hinomoto-family-3b-v3,
  title  = {HinoMoto-Family-3B-v3: QLoRA SFT for Japanese family conversation},
  author = {{Project HinoMoto}},
  year   = {2026},
  url    = {https://huggingface.co/FiShota/hinomoto-family-3b-v3},
}

Acknowledgements

Base: SB Intuitions Sarashina2.2-3B-instruct
HinoMotoBench-ja v0.5 (own)
HuggingFace ecosystem, peft, bitsandbytes, transformers