LocFT-BF

This repository hosts the code and data for the paper: Fine-tuning Done Right in Model Editing.

Requirements:

Environment: requirements.txt (Please use Python 3.9+ for this repository)
```
pip install -r requirements.txt
```
Datasets: The processed data from ZsRE, COUNTERFACT, and WikiBigEdit used in our paper are organized and provided in the ./data/ directory.
Configurations: We have provided default configurations for fine-tuning various LLMs in the ./hparams/ directory (e.g., ./hparams/qwen2.5-7b.yaml). You can modify these configuration files to specify the model path, device ID, or hyperparameters as needed for fine-tuning.

⚠️ Important Note on the EOS Token

Unlike previous model editing codebases, we explicitly append an end-of-sequence (EOS) token to all target answers.

Why? While previous model editing implementations omit the EOS token, typically causing irrelevant or incorrect generation after the correct answer, standard finetuning practices strictly require it (e.g., LlamaFactory and verl).
Impact: This ensures a fair and rigorous comparison and aligning model editing with mainstream LLM training practices.

Editing/Fine-tuning:

To run the individual editing/fine-tuning procedure, you only need to specify the path to the configuration file.

python fine-tune.py --config_path ./hparams/qwen2.5-7b.yaml

Evaluation

Editing Metric Evaluation

To evaluate editing metrics reliability (editing success rate) and generalization (success rate for rephrased prompts), you can run:

CUDA_VISIBLE_DEVICES=0 python eval_edit_metric.py --data_path ./data/zsre/zsre_eval_3k.json --model_path ./saves/qwen --tp_size 1 --num_samples 3000

General Task Evaluation

To evaluate the general capabilities of edited LLMs, we recommend using the lm-evaluation-harness for comprehensive assessment. Additionally, we plan to support representative task evaluation based on vllm in the near future.

Reproducing Main Results

To seamlessly reproduce the main experimental results from our paper (Table 3), which evaluate across three datasets and three LLMs, we provide an automated shell script.

chmod +x main_exp.sh
./main_exp.sh 2>&1 | tee main_exp_logs.txt

Citation

If you have any questions, please feel free to open an issue or contact us. And if you find our work helpful, please cite our paper~

@inproceedings{
yang2026finetuning,
title={Fine-tuning Done Right in Model Editing},
author={Wanli Yang and Rui Tang and Hongyu Zang and Du Su and Qi Cao and Jingang Wang and Huawei Shen and Xueqi Cheng and Fei Sun},
booktitle={The Fourteenth International Conference on Learning Representations},
year={2026},
url={https://openreview.net/forum?id=cfHuA5jsPt}
}

WanliYoung/FT4Editing