Llama 3 Youko 8B Instruct (rinna/llama-3-youko-8b-instruct)

Overview

The model is the instruction-tuned version of rinna/llama-3-youko-8b, using supervised fine-tuning (SFT), Chat Vector, and direct preference optimization (DPO). It adpots the Llama-3 chat format.

• Model architecture

  A 32-layer, 4096-hidden-size transformer-based language model. Refer to the Llama 3 Model Card for architecture details.

• Training: Built with Meta Llama 3

  Supervised fine-tuning. The supervised fine-tuning data is a subset of the following datasets.

  • CohereForAI/aya_dataset
    • The JPN subset was used.
  • FLAN
  • kunishou/databricks-dolly-15k-ja
  • kunishou/hh-rlhf-49k-ja
  • kunishou/oasst1-chat-44k-ja
  • kunishou/oasst2-chat-68k-ja
  • meta-math/MetaMathQA
    • The following sections were used: MATH_AnsAug, MATH_Rephrased, MATH_SV, and MATH_FOBAR.
    • The remaining sections, containing augmented data from commonly used evaluation corpora, were skipped for preventing any possibility of data leak.
  • OpenAssistant/oasst1
    • The EN and JA subsets were used.
  • OpenAssistant/oasst2
    • The EN and JA subsets were used.
  • sahil2801/CodeAlpaca-20k
  • rinna Dataset

  Model merging. The fine-tuned model (llama-3-youko-8b-sft) has been enhanced through the following chat vector addition. The chat vector was obtained by subtracting the parameter vectors of meta-llama/Meta-Llama-3-8B from those of meta-llama/Meta-Llama-3-8B-Instruct.

  llama-3-youko-8b-sft + 0.5 * (meta-llama/Meta-Llama-3-8B-Instruct - meta-llama/Meta-Llama-3-8B)
  

  Here, the embedding layer was skipped while subtracting and adding the parameter vectors.

  Direct preference optimization was then applied with a subset of the following datasets to build this instruct model.

  • kunishou/HelpSteer-35k-ja
  • kunishou/HelpSteer2-20k-ja
  • rinna Dataset

• Contributors

  • Xinqi Chen
  • Koh Mitsuda
  • Toshiaki Wakatsuki
  • Kei Sawada

• Release date

  July 25, 2024

Benchmarking

Please refer to rinna's LM benchmark page (Sheet 20240725).

How to use the model

We found this instruction-tuned model tends to generate repeated text more often than its base counterpart, and thus we set repetition_penalty=1.1 for better generation performance. The same repetition penalty was applied to the instruction-tuned model in the aforementioned evaluation experiments.

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "rinna/llama-3-youko-8b-instruct"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
	model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

messages = [
    {"role": "system", "content": "あなたは誠実で優秀なアシスタントです。どうか、簡潔かつ正直に答えてください。"},
    {"role": "user", "content": "西田幾多郎とはどんな人物ですか？"},
]

input_ids = tokenizer.apply_chat_template(
	messages,
    add_generation_prompt=True,
    return_tensors="pt"
).to(model.device)

terminators = [
	tokenizer.convert_tokens_to_ids("<|end_of_text|>"),
    tokenizer.convert_tokens_to_ids("<|eot_id|>")
]

outputs = model.generate(
	input_ids,
    max_new_tokens=512,
    eos_token_id=terminators,
    do_sample=True,
    temperature=0.6,
    top_p=0.9,
    repetition_penalty=1.1,
)
    
response = outputs[0][input_ids.shape[-1]:]
response = tokenizer.decode(response, skip_special_tokens=True)
print(response)

Tokenization

The model uses the original meta-llama/Meta-Llama-3-8B-Instruct tokenizer.

How to cite

@misc{rinna-llama-3-youko-8b-instruct,
    title = {rinna/llama-3-youko-8b-instruct},
    author = {Chen, Xinqi and Mitsuda, Koh and Wakatsuki, Toshiaki and Sawada, Kei},
    url = {https://huggingface.co/rinna/llama-3-youko-8b-instruct}
}

@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}}
}

References

@article{llama3modelcard,
    title = {Llama 3 Model Card},
    author = {AI@Meta},
    year = {2024},
    url = {https://github.com/meta-llama/llama3/blob/main/MODEL_CARD.md}
}

@article{huang2023chat,
    title = {Chat Vector: A Simple Approach to Equip LLMs with Instruction Following and Model Alignment in New Languages},
    author = {Huang, Shih-Cheng and Li, Pin-Zu and Hsu, Yu-Chi and Chen, Kuang-Ming and Lin, Yu Tung and Hsiao, Shih-Kai and Tzong-Han Tsai, Richard and Lee, Hung-yi},
    year = {2023},
    url = {https://arxiv.org/abs/2310.04799}
}

License

Meta Llama 3 Community License