Official code implementation of "EAMET: ROBUST MASSIVE MODEL EDITING VIA EMBEDDING ALIGNMENT OPTIMIZATION (https://arxiv.org/abs/2505.11876)"

• [2026.01.26] Our work is accepted to ICLR'26!

At least one NVIDIA GPU with 80GB

• python==3.9.21
• torch==2.0.0
• tokenizers==0.21.0
• torchaudio==2.0.0
• torchvision==0.15.0
• transformers==4.49.0
• datasets==1.18.0
• nltk==3.6.5
• numpy==1.22.4
• pandas==2.2.3
• scipy==1.13.1
• scikit-learn==1.6.1
• matplotlib==3.5.1

The main configuration is done through general.sh. Here are the key parameters you can customize:

1. Algorithm Selection (alg_name):

   • EAMET (default)
   • MEMIT
   • PMET
   • ROME
   • FT
   • MEND
   • ALPHAEDIT

2. Model Selection (model_name):

   • NousResearch/Llama-2-7b-hf (default)
   • meta-llama/Llama-3.1-8B
   • NousResearch/Llama-2-13b-hf
   • tiiuae/falcon-7b
   • deepseek-ai/deepseek-llm-7b-base
   • Qwen/Qwen2.5-7B
   • google/gemma-7b-it
   • microsoft/phi-1_5

3. Dataset Selection (ds_name):

   • counterfact (default)
   • zsre
   • wikirecent

4. Hyperparameters:

   • Choose appropriate hparams_fname based on your model.

5. GLEU Benchmark Evaluation: To evaluate the edited models using the GLEU benchmark, modify the evaluation command in general.sh:

   - python -m experiments.evaluate \
   + python -m experiments.evaluate_gleu \

   The GLEU benchmark provides additional metrics for assessing the general ability of edited models.

1. Configure your parameters in general.sh
2. Run the script:

   bash general.sh

• Set dataset_size_limit to control the number of editing tasks (default: 10000)
• Use --use_cache flag to cache KV pairs if needed
• Adjust assigned_prefix_len for evaluation (default: 5)

Results will be saved in the specified output directory with your chosen ./results/out_name.