README.md

Data Cleaning Scripts

Prerequisites

• Yosys
• Icarus Verilog
• Cadence Genus (If you want to use it as your synthesis tool)

Download Datasets

Run the following command to download the RTLCoder and Verigen dataset and write both datasets to the output_dir:

python data/download.py --rtlcoder data/dataset/raw/Resyn27k.json --output_dir data/dataset/raw/verilog 

Then run deduplicate to remove duplicates from the raw Verilog dataset:

python data/deduplicate.py --data data/dataset/raw/verilog

Seperate the Dataset into Clean and Unclean designs

Then, we need to separete the dataset into clean and unclean designs (to be fixed later):

python data/cleaner.py --data dataset/raw/verilog --clean_dir dataset/clean --leftover_dir dataset/leftover --check yosys

Fix the Erroneous Designs Using GPT/Gemini

First, set your Gemini or OpenAI keys depending on which LLM you want to use for fixing the designs:

export GEMINI_KEY=<your-gemini-key>
export OPENAI_KEY=<your-openai-key>

Then, run the auto-fix flow on the unclean designs using your GPT/Gemini:

python data/fixer.py --input_dir --input_dir data/dataset/leftover --output_dir data/dataset/clean --llm gpt --check yosys

Check Syntax and Synthesis Errors of the Clean Set

Make sure that your clean set is synthesizable and error free:

python data/checker.py --data data/dataset/clean --check yosys