README.md

LaM4Inv

A loop invariant inference tool which combines LLMs with traditional Bounded Model Checking tools..

Install

Our current tool only runs on windows system.

1. cd /LM2C/
2. If you want to run experiments on GPT models, prepare a python environment with python version = 3.7.10, then run the command pip install -r requirements.txt. You should input your own API key in GPT_chat/GPT.py. You should disable all from GPT_chat import Llama3chat in RunAllLinear.py and GPT_chat/GPT.py.
3. If you want to run experiments on Llama3-8B, prepare a python environment with python version >= 3.8, then run the command pip install -r requirementsllama.txt. The recommended configuration is a GPU higher than 4090 and more than 48G of RAM.

Run

1. cd /LM2C/
2. All parameter configurations are in the Config.py file and you can modify them as needed.
3. Before you start experimenting, you have to go to the Result folder and create the output folder and .txt file according to the parameter in Config.py. If you want to run all linear benchmarks, python RunAllLinear.py.
4. The results can be founded in Result directory.
5. If you want to run specific one benchmark, change the range of i in RunAllLinear.py.

###Analyze

1. All result are stored in Result.
2. python averageTimeAndProposal.py If you want to analyze the results of the experiment and output the average time spent and the number of proposals.

Benchmarks

1. All benchmarks are put in "Benchmarks/", each instance has three files: c source file, CFG json file, and SMT file.
2. If you want to add new instance, you only need to prepare the three files.
3. As for how to prepare the CFG json file and SMT file, please refer to Code2Inv, which use Clang to do it automatically.