sbintuitions/sarashina2.2-0.5b-instruct-v0.1

Model Summary

This repository provides Japanese language models trained by SB Intuitions.

Model Details

• Model type: Autoregressive Language Model
• Language(s): Japanese

Evaluation in Japanese and English Tasks

Model	Elyza-tasks-100	Japanese MT Bench	English MT Bench
Qwen/Qwen2.5-0.5B-instruct	1.53	2.95	4.98
sarashina2.2-0.5B-instruct-v0.1	2.38	4.55	5.09
Rakuten/RakutenAI-2.0-mini-instruct	2.41	4.49	5.13
SakanaAI/TinySwallow-1.5B-Instruct	2.81	5.24	6.31
Qwen/Qwen2.5-1.5B-instruct	2.28	4.06	6.99
llm-jp/llm-jp-3-1.8b-instruct3	2.53	4.62	4.83
sarashina2.2-1B-instruct-v0.1	2.88	5.09	6.46
google/gemma-2-2b-jpn-it	3.02	5.19	7.56
Qwen/Qwen2.5-3B-instruct	2.99	5.68	7.88
llm-jp/llm-jp-3-3.7b-instruct3	2.79	4.98	5.44
sarashina2.2-3B-instruct-v0.1	3.75	6.51	7.71

How to Use

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline, set_seed

# モデルのロード
model_name = "sbintuitions/sarashina2.2-0.5b-instruct-v0.1"
model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.bfloat16, device_map="auto")
tokenizer = AutoTokenizer.from_pretrained(model_name)
chat_pipeline = pipeline("text-generation", model=model, tokenizer=tokenizer)
set_seed(123)

# ユーザーの入力
user_input = [{"role": "user", "content": "こんにちは。あなたの名前を教えて"}]

# モデルによる応答生成
responses = chat_pipeline(
    user_input,
    max_length=50,
    do_sample=True,
    num_return_sequences=3,
)

# 応答を表示
for i, response in enumerate(responses, 1):
    print(f"Response {i}: {response['generated_text']}")

# Response 1: [{'role': 'user', 'content': 'こんにちは。あなたの名前を教えて'}, {'role': 'assistant', 'content': 'Sarashina2と言います。本日のご要件を教えて下さい。'}]
# Response 2: [{'role': 'user', 'content': 'こんにちは。あなたの名前を教えて'}, {'role': 'assistant', 'content': 'こんにちは！私の名前はSarashina2です。今日はどうしましたか？'}]
# Response 3: [{'role': 'user', 'content': 'こんにちは。あなたの名前を教えて'}, {'role': 'assistant', 'content': 'Sarashina2と言います。本日のご要件を教えて下さい。'}]

Limitations

This model has limited safety training. Therefore, it might generate some meaningless sequences, some inaccurate instances, or biased/objectionable outputs. Before using it, we would like developers to tune models based on human preferences and safety considerations.

License

MIT License