metadata

language:
  - ja
  - en
license: llama2
datasets:
  - databricks/databricks-dolly-15k
  - kunishou/databricks-dolly-15k-ja
  - izumi-lab/llm-japanese-dataset
thumbnail: https://github.com/rinnakk/japanese-pretrained-models/blob/master/rinna.png
inference: false
model-index:
  - name: youri-7b-chat
    results:
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: AI2 Reasoning Challenge (25-Shot)
          type: ai2_arc
          config: ARC-Challenge
          split: test
          args:
            num_few_shot: 25
        metrics:
          - type: acc_norm
            value: 51.19
            name: normalized accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=rinna/youri-7b-chat
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: HellaSwag (10-Shot)
          type: hellaswag
          split: validation
          args:
            num_few_shot: 10
        metrics:
          - type: acc_norm
            value: 76.09
            name: normalized accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=rinna/youri-7b-chat
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: MMLU (5-Shot)
          type: cais/mmlu
          config: all
          split: test
          args:
            num_few_shot: 5
        metrics:
          - type: acc
            value: 46.06
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=rinna/youri-7b-chat
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: TruthfulQA (0-shot)
          type: truthful_qa
          config: multiple_choice
          split: validation
          args:
            num_few_shot: 0
        metrics:
          - type: mc2
            value: 41.17
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=rinna/youri-7b-chat
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: Winogrande (5-shot)
          type: winogrande
          config: winogrande_xl
          split: validation
          args:
            num_few_shot: 5
        metrics:
          - type: acc
            value: 75.06
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=rinna/youri-7b-chat
          name: Open LLM Leaderboard
      - task:
          type: text-generation
          name: Text Generation
        dataset:
          name: GSM8k (5-shot)
          type: gsm8k
          config: main
          split: test
          args:
            num_few_shot: 5
        metrics:
          - type: acc
            value: 1.52
            name: accuracy
        source:
          url: >-
            https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=rinna/youri-7b-chat
          name: Open LLM Leaderboard
base_model: rinna/youri-7b

`rinna/youri-7b-chat`

Overview

The model is the instruction-tuned version of rinna/youri-7b. It adopts a chat-style input format.

Model architecture

A 32-layer, 4096-hidden-size transformer-based language model. Refer to the llama2 paper for architecture details.
Fine-tuning

The fine-tuning data is the subset of the following datasets.
- Databricks Dolly data
- Japanese Databricks Dolly data
- Anthropic HH RLHF data and its Japanese translation
- FLAN Instruction Tuning data and its Japanese translation
- Izumi lab LLM Japanese dataset
  - The following sections are used
    - alt
    - aozora-txt
    - CourseraParallel
    - ParaNatCom
    - Tab-delimited_Bilingual_Sentence_Pairs
    - tanaka-corpus
    - wikinews
    - wordnet
    - yasashi-japanese
  - The remaining sections contain commonly used evaluation corpora so they are skipped to prevent data leak.
Contributors
- Tianyu Zhao
- Kei Sawada
Release date

October 31, 2023

Benchmarking

Please refer to rinna's LM benchmark page (Sheet 20231031).

How to use the model

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("rinna/youri-7b-chat")
model = AutoModelForCausalLM.from_pretrained("rinna/youri-7b-chat")

if torch.cuda.is_available():
    model = model.to("cuda")

instruction = "次の日本語を英語に翻訳してください。"
input = "自然言語による指示に基づきタスクが解けるよう学習させることを Instruction tuning と呼びます。"

context = [
    {
        "speaker": "設定",
        "text": instruction
    },
    {
        "speaker": "ユーザー",
        "text": input
    }
]
prompt = [
    f"{uttr['speaker']}: {uttr['text']}"
    for uttr in context
]
prompt = "\n".join(prompt)
prompt = (
    prompt
    + "\n"
    + "システム: "
)
token_ids = tokenizer.encode(prompt, add_special_tokens=False, return_tensors="pt")

with torch.no_grad():
    output_ids = model.generate(
        token_ids.to(model.device),
        max_new_tokens=200,
        do_sample=True,
        temperature=0.5,
        pad_token_id=tokenizer.pad_token_id,
        bos_token_id=tokenizer.bos_token_id,
        eos_token_id=tokenizer.eos_token_id
    )

output = tokenizer.decode(output_ids.tolist()[0])
print(output)
"""
設定: 次の日本語を英語に翻訳してください。
ユーザー: 自然言語による指示に基づきタスクが解けるよう学習させることを Instruction tuning と呼びます。
システム:  Learning to solve tasks based on natural language instructions is called instruction tuning.</s>
"""

output = output[len(prompt):-len("</s>")].strip()
input = "大規模言語モデル（だいきぼげんごモデル、英: large language model、LLM）は、多数のパラメータ（数千万から数十億）を持つ人工ニューラルネットワークで構成されるコンピュータ言語モデルで、膨大なラベルなしテキストを使用して自己教師あり学習または半教師あり学習によって訓練が行われる。"

context.extend([
    {
        "speaker": "システム",
        "text": output
    },
    {
        "speaker": "ユーザー",
        "text": input
    }
])
prompt = [
    f"{uttr['speaker']}: {uttr['text']}"
    for uttr in context
]
prompt = "\n".join(prompt)
prompt = (
    prompt
    + "\n"
    + "システム: "
)
token_ids = tokenizer.encode(prompt, add_special_tokens=False, return_tensors="pt")

with torch.no_grad():
    output_ids = model.generate(
        token_ids.to(model.device),
        max_new_tokens=200,
        do_sample=True,
        temperature=0.5,
        pad_token_id=tokenizer.pad_token_id,
        bos_token_id=tokenizer.bos_token_id,
        eos_token_id=tokenizer.eos_token_id
    )

output = tokenizer.decode(output_ids.tolist()[0])
print(output)
"""
設定: 次の日本語を英語に翻訳してください。
ユーザー: 自然言語による指示に基づきタスクが解けるよう学習させることを Instruction tuning と呼びます。
システム: Learning to solve tasks based on natural language instructions is called instruction tuning.
ユーザー: 大規模言語モデル（だいきぼげんごモデル、英: large language model、LLM）は、多数のパラメータ（数千万から数十億）を持つ人工ニューラルネットワークで構成されるコンピュータ言語モデルで、膨大なラベルなしテ キストを使用して自己教師あり学習または半教師あり学習によって訓練が行われる。
システム:  Large language models (LLMs) are computer language models consisting of a deep artificial neural network with millions to billions of parameters that are trained by self-supervised learning or semi-supervised learning using vast unlabeled text corpora.</s>
"""

Tokenization

The model uses the original llama-2 tokenizer.

How to cite

@misc{rinna-youri-7b-chat,
    title = {rinna/youri-7b-chat},
    author = {Zhao, Tianyu and Sawada, Kei},
    url = {https://huggingface.co/rinna/youri-7b-chat}
}

@inproceedings{sawada2024release,
    title = {Release of Pre-Trained Models for the {J}apanese Language},
    author = {Sawada, Kei and Zhao, Tianyu and Shing, Makoto and Mitsui, Kentaro and Kaga, Akio and Hono, Yukiya and Wakatsuki, Toshiaki and Mitsuda, Koh},
    booktitle = {Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024)},
    month = {5},
    year = {2024},
    pages = {13898--13905},
    url = {https://aclanthology.org/2024.lrec-main.1213},
    note = {\url{https://arxiv.org/abs/2404.01657}}
}

License

The llama2 license

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	48.51
AI2 Reasoning Challenge (25-Shot)	51.19
HellaSwag (10-Shot)	76.09
MMLU (5-Shot)	46.06
TruthfulQA (0-shot)	41.17
Winogrande (5-shot)	75.06
GSM8k (5-shot)	1.52