rinna-neox-small-ja-it-adapter

LoRA adapter for instruction-tuned rinna/japanese-gpt-neox-small in Japanese.

• Task: instruction-following / text generation
• Language: Japanese
• License: CC BY-SA 4.0
• Base model: rinna/japanese-gpt-neox-small
• Release type: LoRA adapter weights (base model required)

Usage

Load the base model and apply the LoRA adapter:

from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel
import torch

base_id = "rinna/japanese-gpt-neox-small"
adapter_id = "takehika/rinna-neox-small-ja-it-adapter"
tokenizer = AutoTokenizer.from_pretrained(base_id, use_fast=False)
if tokenizer.pad_token is None:
    tokenizer.pad_token = tokenizer.eos_token
base_model = AutoModelForCausalLM.from_pretrained(base_id)
model = PeftModel.from_pretrained(base_model, adapter_id).eval()

def build_prompt(instruction, input_text=""):
    if input_text:
        return (
            "### Instruction:\n"
            f"{instruction}\n\n"
            "### Input:\n"
            f"{input_text}\n\n"
            "### Response:\n"
        )
    else:
        return (
            "### Instruction:\n"
            f"{instruction}\n\n"
            "### Response:\n"
        )


instruction = "毎日健康的に過ごすコツを5つ挙げてください。"
prompt = build_prompt(instruction)

inputs = tokenizer(prompt, return_tensors="pt", add_special_tokens=False)
with torch.no_grad():
    output_ids = model.generate(
        **inputs,
        max_new_tokens=128,
        do_sample=True,
        pad_token_id=tokenizer.eos_token_id,
        eos_token_id=tokenizer.eos_token_id,
    )

gen_ids = output_ids[0][inputs["input_ids"].shape[1]:]
generated = tokenizer.decode(gen_ids, skip_special_tokens=True)
print(generated)

Data

• Dataset:
  • llm-jp/llm-jp-instructions
  • kunishou/databricks-dolly-15k-ja

Training

This adapter is instruction-tuned with a prompt-response format:

### Instruction:
{instruction}

### Input:
{input}

### Response:
{response}

LoRA adapters are trained on the base model, and the adapter is applied at inference time by loading the base model and the adapter weights.

Intended Use & Limitations

• Intended for Japanese instruction-following generation.
• Outputs may be verbose or partially off-instruction.
• It can produce incorrect or misleading content; verify critical outputs.

Attribution & Licenses

• License: CC BY-SA 4.0
• Base model: rinna/japanese-gpt-neox-small — MIT License
  • Model card: https://huggingface.co/rinna/japanese-gpt-neox-small
• Dataset: llm-jp/llm-jp-instructions — CC BY 4.0
  • Dataset card: https://huggingface.co/datasets/llm-jp/llm-jp-instructions
• Dataset: kunishou/databricks-dolly-15k-ja — CC BY-SA 3.0
  • Dataset card: https://huggingface.co/datasets/kunishou/databricks-dolly-15k-ja

This adapter modifies the base model by fine-tuning on the above datasets.

Base Model Citation

@misc{rinna-japanese-gpt-neox-small,
    title = {rinna/japanese-gpt-neox-small},
    author = {Zhao, Tianyu and Sawada, Kei},
    url = {https://huggingface.co/rinna/japanese-gpt-neox-small}
}

@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}}
}