CharacterBot

Overview

We introduce CharacterBot, a model designed to replicate both the linguistic patterns and distinctive thought processes of a character. In this repository, we present Lu Xun—a renowned Chinese writer—as a case study.

Dataset

The experimental dataset luxun_essay.json comprises 638 articles sourced from Wikisource. These collections span the entirety of Lu Xun’s mature essayistic output, reflecting a diverse array of themes from his intellectual career. All texts are in the public domain, ensuring unrestricted scholarly usage. As Lu Xun was a Chinese writer, the repository includes Chinese textual data and code annotations.

Requirements

To set up the environment and install necessary dependencies, follow the steps below:

• python3
• conda create --name env
• pip3 install -r requirements.txt

How to Use

Data Generation

Authorial Perspective Reframing

Generate the Authorial Perspective Reframing pre-training data for the following commands:

cd characterbot
python authorial_perspective_reframing.py

Multiple-choice Questions

Generate data for the multiple-choice questions task:

python multiple_choice_questions_data.py

Generative Question Answering

Generate data for the generative question answering task:

python generative_qa_data.py

Style Transfer

Generate data for the style transfer task:

python style_transfer_data.py

Training

CharLoRA

CharLoRA requires modifications to three editable Python libraries. Make sure to download these libraries as specified in the project requirements. Then, apply the following updates:

Modifications for LLaMA-Factory:

cd /path/to/characterbot/train_with_charlora
cp aligner.py /path/to/LLaMA-Factory/src/llamafactory/data/aligner.py
cp collator.py /path/to/LLaMA-Factory/src/llamafactory/data/collator.py
cp supervised.py /path/to/LLaMA-Factory/src/llamafactory/data/processors/supervised.py

Modifications for PEFT:

cp peft_model.py /path/to/peft/src/peft/peft_model.py
cp save_and_load.py /path/to/peft/src/peft/utils/save_and_load.py
cp layer.py /path/to/peft/src/peft/tuners/lora/layer.py

Modifications for Transformers:

cp modeling_qwen2.py /path/to/transformers/src/transformers/models/qwen2/modeling_qwen2.py

Data Integration with LLaMA-Factory

Since both pre-training and fine-tuning processes rely on LLaMA-Factory, integrate your generated data into its structure.

Place data files into LLaMA-Factory's data directory:

cp pre_train.json /path/to/LLaMA-Factory/data/pre_train.json
cp fine_tune.json /path/to/LLaMA-Factory/data/fine_tune.json
cp test.json /path/to/LLaMA-Factory/data/test.json
cp dataset_info.json /path/to/LLaMA-Factory/data/dataset_info.json

Place training configuration files into the examples directory:

cp pre_train.yaml /path/to/LLaMA-Factory/examples/train_lora/pre_train.yaml
cp fine_tune.yaml /path/to/LLaMA-Factory/examples/train_lora/fine_tune.yaml

Pre-training

Run pre-training using the following command:

cd /path/to/LLaMA-Factory
llamafactory-cli train examples/train_lora/pre_train.yaml

Fine-tuning

Run fine-tuning with this command:

llamafactory-cli train examples/train_lora/fine-tuning.yaml

The output model will be saved at /path/to/LLaMA-Factory/saves/fine_tune

Testing

Generating Responses

After training, you can generate responses for each task. Make sure to update the model and data paths as needed. Some data example format are shown in example. For each task, follow these instructions:

Multiple-choice Questions:

cd /path/to/characterbot/test/
cp peft_model_test.py /path/to/peft/src/peft/peft_model.py
cp mcq_load.py /path/to/peft/src/peft/utils/save_and_load.py
python multiple_choice_gen.py

Generative Question Answering:

cp gqa_load.py /path/to/peft/src/peft/utils/save_and_load.py
python generate_qa_gen.py

Style Transfer:

cp st_load.py /path/to/peft/src/peft/utils/save_and_load.py
python style_transfer_gen.py

Evaluation

Evaluate the performance of the model. Note that the evaluation code for multiple-choice questions is the same as the generation code. For the remaining tasks, use the following commands:

Generative Question Answering Evaluation:

python gen_qa_eval.py

Style Transfer Evaluation:

cd /path/to/characterbot/test/style_transfer_eval
python bleu.py
python rouge.py
python style_matching.py

Acknowledgements

Parts of this project is built upon the following open-source libraries:

• transformers (Apache 2.0)
• peft (Apache 2.0)
• LLaMA-Factory (Apache 2.0)