How to discuss about the project? Is their any discord channel?

[thePromger]

Hi
i really liked this project not to just decompile files but the way it works
Actually i didn't go through it till yet, but i did some testing on your hosted version

and i saw mostly are indentation errors which do create errors and LLM tries to add wrong code, which breaks it completely
so i also want to learn and help in improving it
do let me know if their is any discord, coz discussing their will be much easier

Suggestions & Problems :

• try to break down the process per class/method/function and then concatenate them in single file according to the order
  it way chances becomes else that the suggestion by LLM breaks something else

• we even need to add some logic too about indentation, try&except, if & else rules

• whenever LLM saw any if, it tries to add else to patch the error but it doesn't needed their at all, the actually reason was the indentation error mostly & LLM just adds if or else mostly which adds more bloat/errors to code, even python interpreter also gives good suggestions to fix the error like identation, if or else, try & except, etc

[jikk]

@thePromger thanks for your suggestion. @jdw170000 @Joel-Flores123 Would you respond?

[jdw170000]

We currently don't have the resources to moderate a public Discord server, so GitHub issues are the primary forum for discussing problems and solutions for the decompiler. We would be interested in getting in contact with people who have experience running community discords to explore our options there.

Regarding the suggestions, I would like to clarify a couple things about how PyLingual functions under the hood:

• PyLingual already processes each code object (classes, functions, lambdas, etc.) independently, and those results are combined together by the control flow reconstructor. The two language models in PyLingual are the segmentation model, which chunks bytecode instructions into statements, and the statement translator, which independently translates single statements of bytecode into single statements of source code. These language models do not see or respond to errors in the decompilation.
• The statement translation model outputs a flat list of source code statements, which are then combined and indented by the control flow reconstructor, which is already a rule-based system. Any else: blocks are inserted by the control flow reconstructor. We are working to improve the control flow pattern matching in the dev branch. If you have some simplified cases that demonstrate the indentation errors you're referring to, then we can add those to our test suite and improve the control flow reconstruction accuracy.

The version on the web service is using a legacy version of the control flow reconstructor that currently has marginally better performance on older Python versions, but we plan to replace it with the open source version later this year as we continue to make improvements here.

[thePromger]

@jdw170000

I can help you in running discord community

Is their any pdf or something to read more about how pylingual works?
Can I also join the team for this project?
I tried research paper but unable to understand it
Like where to start, where it's end, how to read it correctly, etc

[jdw170000]

While the presentation skims over most of the details, we did present an overview of PyLingual at BlackHat USA in 2024 (https://www.youtube.com/watch?v=TJ65LrInHxw), which may be more approachable than the paper.

You are free to make a fork to work on features and improvements, then make a merge request when it's ready for review.

[thePromger]

@jdw170000

What about discord server?
I am new to this, so I am gonna need your help to understand it in better way

[jdw170000]

We are still exploring our options regarding a potential discord server.

I can answer specific questions about the code and architecture of PyLingual, but general conceptual questions would be better answered by the paper or the talks. This one should provide more background about Python bytecode and the interpreter: https://www.youtube.com/watch?v=rZ_YAm9ywuo.