CodeBridge

Replication package for Zero-Shot Cross-Domain Code Search without Fine-Tuning, FSE 2025.

Project Structure

-data # datasets and generated data
    -cross-domain 
    -rapid
-models # base models
-scripts # scripts for data preprocessing and evaluation

Model

We use the following models:

• CoCoSoDa
• UniXcoder
• bge-large-en-v1.5

Environment

• python==3.8.18
• torch==2.0.1
• You can install the dependencies by:

pip install -r requirements.txt

Evaluation in Cross-Domain Setting

• We have provided the generated files in data/. For instructions on generating the code and comments using DeepSeek-Coder-1.3b-Instruct, please refer to generate_code.py and generate_comment.py.
• To reproduce the result, start by obtaining the embeddings from various models. The embeddings will be stored in the folder vectors. You can modify the dataset name in scripts/get_embedding.sh to get different embeddings.

bash scripts/get_embedding.sh

• Evaluate the results.

bash scripts/eval.sh

Evaluation in RAPID Setting

• Get embeddings.

bash scripts/get_embedding_rapid.sh

• Evaluate the results.

bash scripts/eval_rapid.sh