albertan017

> Are you training on a single node or multiple nodes out of interest? For the 1B model, we use a single node. For larger models, they are typically trained...

Thanks for your interest! Integration with Ghidra and IDA Pro is definitely on our roadmap. Currently, we are concentrating on training a new large language model designed for binary analysis....

We've only found AnghaBench and Exebench, which cover nearly all available C libraries. If you have specific requirements, you might need to manually compile larger projects like Linux. While it's...

可以参考下面的issue https://github.com/albertan017/LLM4Decompile/issues/33 目前没有统一的评估方式，我们也在探索不同的评估方法，比如使用gpt评估： https://github.com/albertan017/LLM4Decompile/blob/main/samples/readability_template.txt 如果有其他推荐的评估方式，欢迎讨论交流~

For compilable data, you may follow the [compilation script for AnghaBench](https://github.com/albertan017/LLM4Decompile/blob/main/train/compile.py), with small modification on handling the source of function (exebench_data['func_def']) and its dependency (exebench_data['synth_deps']). For executable data, it's quite...

Yes, in theory, it should be effective. However, we encounter difficulties in generating the appropriate assembly for execution. As a result, we adjust the input to the Wrapper and alter...

As highlighted in our paper, we initially eliminate functions that cannot be executed by testing the executability of the original function (i.e., **use the dataset_row['func_def']**, not the 'decompiled_c_func' in step...

目前的llm并不具备项目级代码理解能力（llm翻译一段话很简单，翻译一个章节明显出现遗忘问题），训练和推理开销也是极其高（不考虑优化，attention计算是输入长度的三次方关系），训练项目级重建成本和难度太高。 我们更倾向于单独重构，整合重组：利用好函数自身的信息去重构，再将一个个重构的函数一起送入更强的模型（GPT-o1，Deepseek-R1）去refine。llm4decompile负责做好单个函数，GPT等则擅长从更高层次整合数据

The 9B model is based on [Yi-Coder](https://github.com/01-ai/Yi-Coder), while the training script is from [Deepseek-Coder](https://github.com/deepseek-ai/DeepSeek-Coder). We did not test the 9B model for the script, we recommend to use llama factory...

2024.5.10 Update: All the evaluations and models are based on executable! enjoy~ ~~Thanks for your interest for our project! Indeed, we're utilizing object files instead of executables, as our training...