SelfPiCo: Self-Guided Partial Code Execution with LLMs (ISSTA'24)

How to use

Step 0: Before we start

This project is developed based Lexecutor, so we strongly recommend to read the document of Lexecutor first.

SelfPiCo-GPT

Step 1: Instrument the partial code snippet

python -m SelfDyCo.Instrument --files test.py --iids iids.json

Step 2: Set the mode

1. change the base LLM in Runtime.py
2. Using your own Openai API key in SelfPiCo/pedictor/GPT/ValuePrediector.py and SelfPiCo/pedictor/GPT/TypePrediector.py

Step 3: Guide partial code execution

Direct execute the instrumented partial code

SelfPiCo-CodeLlama

Step 1: Instrument the partial code snippet

python -m SelfDyCo.Instrument --files test.py --iids iids.json

Step 2: Fine-tune the CodeLLama model

1. Collect the variable type data by Lexecutor
2. Running SelfPiCo/pedictor/CodeLLama/PrepareData.py
3. Running SelfPiCo/pedictor/CodeLLama/FineTune.py

Step 3: Guide partial code execution

1. start up the model python -m SelfDyCo.predictor.CodeLLama.CodellamaServer
2. Direct execute the instrumented partial code