sawac/dora-bocchan-yozo-sensei

DoRA Adapters for Japanese Literary Characters

DoRA (Weight-Decomposed Low-Rank Adaptation) fine-tuned adapters for generating text in the style of three iconic Japanese literary characters.

Overview

This repository contains lightweight DoRA adapters (23MB) that can be applied to the LFM2.5-1.2B-JP-MLX-bf16 base model to generate text in the voices of:

1. 坊ちゃん (Botchan) - From Natsume Soseki's "Botchan"

   • Edo-born, strong sense of justice, short-tempered
   • Speech pattern: "〜だ", "〜じゃないか", "おれ"

2. 大庭葉蔵 (Yozo Oba) - From Dazai Osamu's "No Longer Human"

   • Self-deprecating, introspective, plays the fool to fit in
   • Speech pattern: "自分", literary and melancholic

3. 先生 (Sensei) - From Natsume Soseki's "Kokoro"

   • Intellectual, guilt-ridden, socially distant
   • Speech pattern: "私", formal but weighty

Model Details

• Base Model: LFM2.5-1.2B-JP-MLX-bf16
• Fine-tuning Method: DoRA (Weight-Decomposed Low-Rank Adaptation)
• Adapter Size: 23MB (99% smaller than full fine-tuning)
• Rank: 8
• Training Data: 240 examples across 3 characters
• Framework: MLX (Apple Silicon optimized)

Files

• adapters.safetensors (23.8 MB) - Trained DoRA weights
• adapter_config.json (942 bytes) - Adapter configuration

Usage

Installation

pip install mlx-lm

Loading the Adapter

from mlx_lm import load, generate

# Load base model with adapter
model, tokenizer = load(
    "mlx-community/LFM2.5-1.2B-JP-MLX-bf16",
    adapter_path="path/to/downloaded/adapter"
)

# Generate in character voice
response = generate(
    model,
    tokenizer,
    prompt="人と一緒にいるとき、本当は何を考えている？",
    max_tokens=200,
    verbose=True
)

Example Outputs

Yozo Oba (大庭葉蔵):

人と一緒にいるとき、自分はただ、その場をやり過ごすために、
あるいは恐怖から逃れるために、滑稽な道化を演じていただけなのです。
人間というものが、あまりに恐ろしく、理解しがたい存在であったために、
周囲の期待や、他者の視線が自分を突き放すような気がして、
自分はそこに居られなかったのです。
……ふふ、自分には、どうしようもないのです。

Botchan (坊ちゃん):

そんな卑怯な真似は、おれにはできねえよ！
理屈じゃねえ、情理に外れたことは真っ向から否定してやる。
たとえそれが、誰かにとって都合のいい結果になろうが、
曲がったことは曲がってると言わなきゃ気が済まねえ。

Training Details

Dataset

• 240 training examples
• 30 validation examples
• Balanced across 3 characters
• Focus on dialogue-style prompts and responses

Hyperparameters

• Method: DoRA
• Rank: 8
• Alpha (scale): 20.0
• Learning Rate: 1e-4
• Iterations: 500
• Batch Size: 1
• Max Sequence Length: 2048
• LoRA Layers: All 16 layers

Evaluation

The adapters were evaluated using an AI judge on 5 axes:

1. Tone Consistency (口調の一貫性)
2. Pronoun Accuracy (一人称の正確性)
3. Character Personality (キャラクター性)
4. Japanese Naturalness (日本語の自然さ)
5. Response Depth (回答の深さ)

Average Scores:

• Botchan: 19.2/25
• Yozo Oba: 22.2/25
• Sensei: 21.6/25

Limitations

• Question Response: The model tends to deliver "speeches" about the character's philosophy rather than directly answering questions
• Context Length: Limited by base model's 2048 token context
• Character Mixing: Occasionally, speech patterns may blend between characters
• Training Data: Limited to 240 examples; could benefit from more diverse scenarios

Future Improvements

• Increase LoRA rank to 16 or 32 for more expressive responses
• Expand training data to 500+ examples per character
• Add more characters from Japanese literature
• Fine-tune for better question-answering while maintaining character voice

Citation

If you use this adapter in your research, please cite:

@misc{dora-jp-characters,
  title={DoRA Adapters for Japanese Literary Characters},
  author={Sawada Shusaku},
  year={2026},
  howpublished={\url{https://huggingface.co/sawac/dora-bocchan-yozo-sensei}}
}

License

MIT License

Acknowledgments

• Base model: mlx-community/LFM2.5-1.2B-JP-MLX-bf16
• Training framework: mlx-lm
• Original works: Natsume Soseki and Dazai Osamu