Please follow the requirements.txt
file to install the relevant dependencies or run:
pip install -r requirements.txt
Since our method modifies the transformers of huggingface, please make sure to install the same transformers as ours (we use transformers version 4.26.1):
cd transformers
pip install -e .
We used two datasets to fine-tune our model:
We use openai's gpt-3.5-turbo to generate the solution plans data for training.You can get it by running generate_plans_from_LLM\generate_apps_plan.py
:
python generate_apps_plan.py --test_path ../data/apps/train --save_path ../data/apps/train --start 0 --end 5000
Using CodeT5 as an example, you can run train_codet5.py
to finetune the CodeT5 with solution plans:
python train_codet5.py --model codet5-large-ntp-py --save_dir ./models -- train_path ./data/appps/train --tuning_mode plan --clone_pl_head True --epochs 10
You can run generate_codet5.py
to generate codes:
python generate_codet5.py --test_path data/apps/test -- output_path ./outputs/codes --model_path ./model --plan_head True --is_plan False --temperature
You can run generate_codet5_plan.py
to generate solution plans:
python generate_codet5_plan.py --test_path ./data/apps/test --output_path ./outputs/plans --model_path ./model --plan_head True --is_plan True --temperature 0.6
You can run 'generate_code_with_plan.py` to generate code with generated solution plan as prompt:
python generate_code_with_plan.py --test_path ./data/apps/test --output_path ./outputs/codes --model_path ./model --plan_head True --is_plan False
You can run test_one_solution.sh
to evaluate generated codes:
bash test_one_solution.sh