Skip to content

README.md

Latest commit

 

History

History
79 lines (71 loc) · 2.24 KB

README.md

File metadata and controls

79 lines (71 loc) · 2.24 KB

CodePLAN

Installation

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 .

Datasets

We used two datasets to fine-tune our model:

  • APPS: You can download it here.
  • MBPP: This dataset can be found here.

Generating Plans from LLM

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

Finetuning with solution plans

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 \
    --epochs 10

Generating codes with finetuned model

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
    --temperature 0.6

Generating solution plans with finetuned model

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 \
    --is_plan \
    --temperature 0.6

Generating codes with solution plan

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

Evaluate generated codes

You can run test_one_solution.sh to evaluate generated codes:

bash test_one_solution.sh
python eval_metric.py