This repo contains a low-rank adapter for LLaMA-7b fit on the llm-japanese-dataset dataset.

This version of the weights was trained with the following hyperparameters:

• Epochs: 5
• Batch size: 128
• Cutoff length: 256
• Learning rate: 3e-4
• Lora r: 4
• Lora target modules: q_proj, v_proj

import torch
from transformers import LlamaForCausalLM, LlamaTokenizer
from peft import PeftModel

base_model = "decapoda-research/llama-7b-hf"
# Please note that the special license of decapoda-research/llama-7b-hf is applied.
model = LlamaForCausalLM.from_pretrained(base_model, torch_dtype=torch.float16)
tokenizer = LlamaTokenizer.from_pretrained(base_model)
model = PeftModel.from_pretrained(
    model,
    "izumi-lab/llama-7b-japanese-lora-v0",
    torch_dtype=torch.float16,
)

To see more latest information, please go to llm.msuzuki.me.

Details

• Japanese Paper: https://jxiv.jst.go.jp/index.php/jxiv/preprint/view/422
• English Paper:
• GitHub: [https://github.com/retarfi/jallm]
• Website: llm.msuzuki.me.

Citation:

@preprint{Suzuki2023-llmj,
  title={{日本語インストラクションデータを用いた対話可能な日本語大規模言語モデルのLoRAチューニング}},
  author={鈴木 雅弘 and 平野 正徳 and 坂地 泰紀},
  doi={10.51094/jxiv.422},
  archivePrefix={Jxiv},
  year={2023}
}

If you have any inquiries, such as joint research, data provision, various types of support, please email to izumi-llm@socsim.org .