AlphaEdit: Multilingual Model Editing Framework

AlphaEdit is a sophisticated model editing framework designed for multilingual large language models. It implements advanced techniques for editing factual knowledge in transformer-based models while preserving model performance across multiple languages.

AlphaEdit/
├── AlphaEdit/                 # Core AlphaEdit implementation
│   ├── AlphaEdit_main.py     # Main editing algorithm
│   ├── AlphaEdit_hparams.py  # Hyperparameter configuration
│   ├── compute_ks.py         # Key computation utilities
│   └── compute_z.py          # Z computation and fact lookup
├── glue_eval/                # Evaluation suite
│   ├── glue_eval.py         # Main evaluation script
│   ├── sst_eval.py          # Sentiment analysis evaluation
│   ├── xnli_eval.py         # Cross-lingual NLI evaluation
│   ├── mlqa_eval.py         # Multilingual QA evaluation
│   ├── wikiann_eval.py      # Named entity recognition evaluation
│   └── ...                  # Additional evaluation modules
├── dsets/                   # Dataset utilities
├── util/                    # Utility functions
├── hparams/                 # Hyperparameter configurations
└── run.sh                   # Main execution script

• Python 3.8+
• PyTorch 1.12+
• CUDA-compatible GPU (recommended)

pip install torch torchvision transformers
pip install datasets scikit-learn numpy
pip install huggingface-hub tokenizers
pip install ipdb  # for debugging

For evaluation tasks:

pip install jieba  # for Chinese text processing
pip install seqeval  # for sequence labeling evaluation

1. Configure hyperparameters: Edit or create configuration files in hparams/AlphaEdit/

2. Run model editing:

bash run.sh

Key hyperparameters in AlphaEditHyperParams:

• Model Configuration:

  • model_name: Target model identifier
  • layers: List of layers to edit
  • layer_selection: Strategy for layer selection ("all" or "random")

• Editing Parameters:

  • fact_token: Token selection strategy
  • v_num_grad_steps: Number of gradient steps
  • v_lr: Learning rate for value optimization
  • clamp_norm_factor: Normalization factor
  • kl_factor: KL divergence regularization

• Statistics:

  • mom2_dataset: Dataset for moment statistics
  • mom2_n_samples: Number of samples for statistics
  • nullspace_threshold: Threshold for null space projection
  • L2: L2 regularization factor

The framework includes comprehensive evaluation across multiple tasks:

2. Multilingual Tasks:
   • XNLI (Cross-lingual Natural Language Inference)
   • MLQA (Multilingual Question Answering)
   • WikiANN (Multilingual Named Entity Recognition)
   • PAWS-X (Cross-lingual Paraphrase Adversaries)

python3 -m experiments.evaluate \
    --alg_name=AlphaEdit \
    --model_name=meta-llama/Meta-Llama-3-8B-Instruct \
    --hparams_fname=hparams/AlphaEdit/Llama3-8B.json \
    --ds_name=mzsre \
    --dataset_size_limit=800 \
    --num_edits=100

AlphaEdit supports editing in multiple languages:

• English (en): Primary language support
• French (fr): Full editing and evaluation support
• Spanish (es): Comprehensive multilingual editing
• German (de): Cross-lingual knowledge transfer
• Dutch (nl): European language support
• Chinese (zh): Asian language support

• Language-aware context templates
• Multilingual null space projection
• Cross-lingual knowledge preservation
• Language-specific evaluation metrics

This project is licensed under the MIT License - see the LICENSE file for details.

## Acknowledgments

- Built upon the AlphaEdit and MEMIT model editing frameworks