README.md

metadata

license: apache-2.0
language:
  - en
tags:
  - code-generation
  - decompilation
  - cuda
  - ptx
  - llm

Dataset for "Enhancing LLM to Decompile Optimized PTX to Readable CUDA for Tensor Programs" (ASE 2025)

Dataset Description

This dataset accompanies the paper:

Sun X, Tang F, Zhang Y, et al. Enhancing LLM to Decompile Optimized PTX to Readable CUDA for Tensor Programs[C]//2025 40th IEEE/ACM International Conference on Automated Software Engineering (ASE). IEEE, 2025: 2235-2247.

It contains pairs of input PTX kernels and output CUDA kernels generated by two auto-schedulers: Ansor and Welder. The dataset is designed to train and evaluate large language models (LLMs) for the task of decompiling optimized PTX back to human-readable CUDA code.

Dataset Structure

The dataset is split into four Parquet files:

• train_ansor.parquet
• train_welder.parquet
• test_ansor.parquet
• test_welder.parquet

Each file contains samples with the following fields:

Field	Type	Description
file	string	Source JSON filename (without extension). For Ansor, it encodes parameters.
name	string or int	Sample identifier (string for Ansor, integer for Welder).
kernel	string	(Ansor only) Kernel function name.
input	string	Input PTX kernel.
output	string	Output CUDA kernel.

Note: The kernel field is present only in Ansor samples; Welder samples do not have this field.

Data Splits

Split	Number of Samples
train_ansor	371,472
train_welder	383,790
test_ansor	1,905
test_welder	1,793

Usage Example

from datasets import load_dataset

dataset = load_dataset("your-username/ptxdec-dataset")
train_ansor = dataset["train_ansor"]
print(train_ansor[0])