This is the official implementation of the paper:

• Mitigating Non-Representative Prototypes and Representation Bias in Few-Shot Continual Relation Extraction

• [2025\05\16] Our paper has been accepted to ACL 2025 Main Conference!

1. Navigate to the Bert directory:

   cd Bert
   

2. Install necessary packages:

   !pip install transformers==4.40.0 torch==2.3.0 scikit-learn ==1.4.2 nltk==3.8.1 retry==0.9.2
   !pip install flash-attn --no-build-isolation
   !pip install pytorch_metric_learning
   

3. Run the TACRED training:

   python train_tacred_final.py --task_name Tacred --num_k 5 
   

4. Run the FewRel training:

   python train_fewrel_final.py --task_name FewRel --num_k 5 
   

1. Navigate to the Llama2,Llama3,Mistral in 'LLMEs'directory:

2. Install necessary packages:

   !pip install transformers==4.40.0 torch==2.3.0 scikit-learn ==1.4.2 nltk==3.8.1 retry==0.9.2
   !pip install llm2vec==0.2.2
   !pip install flash-attn --no-build-isolation
   !pip install pytorch_metric_learning
   

3. Log in to Hugging Face:

   !huggingface-cli login --token your_huggingface_token_to_access_model
   

1. Run the TACRED training:

   python train_tacred_final.py --task_name Tacred --num_k 5 
   

2. Run the FewRel training:

   python train_fewrel_final.py --task_name FewRel --num_k 5 
   

• CPL results with Llama2 and Mistral are evaluated under float32 for fair comparison with Quyen et al.
• All other experiments use bf16.
• Requires Python ≥3.8.
• For CPL, add your OpenAI API key to config.ini.