sbintuitions/sarashina2.2-3b-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

# ใƒขใƒ‡ใƒซใฎใƒญใƒผใƒ‰
model_name = "sbintuitions/sarashina2.2-3b-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)

# ใƒฆใƒผใ‚ถใƒผใฎๅ…ฅๅŠ›
user_input = [{"role": "user", "content": "ใ“ใ‚“ใซใกใฏใ€‚ใ‚ใชใŸใฎๅๅ‰ใ‚’ๆ•™ใˆใฆ"}]

# ใƒขใƒ‡ใƒซใซใ‚ˆใ‚‹ๅฟœ็ญ”็”Ÿๆˆ
responses = chat_pipeline(
    user_input,
    max_length=50,
    do_sample=True,
    num_return_sequences=3,
)

# ๅฟœ็ญ”ใ‚’่กจ็คบ
for i, response in enumerate(responses, 1):
    print(f"Response {i}: {response['generated_text']}")

# Response 1: [{'role': 'user', 'content': 'ใ“ใ‚“ใซใกใฏใ€‚ใ‚ใชใŸใฎๅๅ‰ใ‚’ๆ•™ใˆใฆ'}, {'role': 'assistant', 'content': 'Sarashina2ใจ่จ€ใ„ใพใ™ใ€‚ๆœฌๆ—ฅใฎใ”่ฆไปถใ‚’ๆ•™ใˆใฆไธ‹ใ•ใ„ใ€‚'}]
# Response 2: [{'role': 'user', 'content': 'ใ“ใ‚“ใซใกใฏใ€‚ใ‚ใชใŸใฎๅๅ‰ใ‚’ๆ•™ใˆใฆ'}, {'role': 'assistant', 'content': 'ใ“ใ‚“ใซใกใฏ๏ผ็งใฎๅๅ‰ใฏSarashina2ใงใ™ใ€‚ไปŠๆ—ฅใฏใฉใ†ใ—ใพใ—ใŸใ‹๏ผŸ'}]
# Response 3: [{'role': 'user', 'content': 'ใ“ใ‚“ใซใกใฏใ€‚ใ‚ใชใŸใฎๅๅ‰ใ‚’ๆ•™ใˆใฆ'}, {'role': 'assistant', 'content': 'Sarashina2ใจ่จ€ใ„ใพใ™ใ€‚ๆœฌๆ—ฅใฎใ”่ฆไปถใ‚’ๆ•™ใˆใฆไธ‹ใ•ใ„ใ€‚'}]

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

Downloads last month
8,088
Safetensors
Model size
3.36B params
Tensor type
BF16
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for sbintuitions/sarashina2.2-3b-instruct-v0.1

Finetuned
(2)
this model
Finetunes
18 models
Quantizations
16 models

Spaces using sbintuitions/sarashina2.2-3b-instruct-v0.1 6

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