---
license: mit
tags:
- Python
- Pytorch
- Triton
- Kernal-Optimization
task_categories:
- text-generation
language:
- en
size_categories:
- 10K<n<100K
---



## Overview

This dataset, `TEEN-D/Code_Opt_Triton`, is an extended version of the publicly available [GPUMODE/Inductor_Created_Data_Permissive](https://huggingface.co/datasets/GPUMODE/Inductor_Created_Data_Permissive) dataset. It contains pairs of original (PyTorch or Triton) programs and their equivalent Triton code (generated by torch inductor), intended for training models in PyTorch-to-Triton code translation and optimization.

The primary modification in this extended version is that each optimized Triton code snippet is paired with both its original source (either the initial PyTorch code or the original Triton code, if available in the source dataset). This effectively doubles the dataset size, providing a richer training ground for models aiming to understand the relationship between different representations of the same GPU kernel.

This dataset was created as part of research exploring the capabilities of Large Language Models (LLMs) in writing and optimizing GPU kernels, motivated by the observation that LLMs often struggle with this task due to potential data scarcity. The goal is to provide a more comprehensive dataset for training models that can bridge the gap between high-level PyTorch code and efficient, low-level Triton kernels.

**License:** This dataset is derived from the [GPUMODE/Inductor_Created_Data_Permissive](https://huggingface.co/datasets/GPUMODE/Inductor_Created_Data_Permissive) dataset, which is released under the **MIT License**. Therefore, this extended version is also shared under the **MIT License**. You are free to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of this dataset, subject to the terms of the MIT License.



## Rationale
This extended dataset, pairing optimized Triton with both original Triton and Python code, enhances training for preference-based methods like GRPO/DPO by:

- Providing richer preference signals through diverse (original, optimized) pairs.
- Improving the LLM's understanding of semantic equivalence between PyTorch and Triton.
- Reducing bias towards a specific input format (Triton vs. Python).
- Potentially creating more meaningful comparison groups for GRPO.
- Offering more diverse "rejected" samples for DPO against the "accepted" optimized Triton.

## Dataset Creation

This dataset was generated by processing the [GPUMODE/Inductor_Created_Data_Permissive](https://huggingface.co/datasets/GPUMODE/Inductor_Created_Data_Permissive) dataset through the following steps:

1.  **Loading the Source Dataset:** The `GPUMODE/Inductor_Created_Data_Permissive` dataset, containing pairs of PyTorch programs and their equivalent Triton code generated by `torch.compile`, was loaded. This original dataset was created by parsing PyTorch programs from publicly available GitHub repositories (based on the Stack v1 dataset), extracting `torch.nn` modules, creating unit tests, and then using `torch.compile` to generate Triton code. The Triton code was further transformed to resemble the format seen in KernelBench, and each entry was enriched with metadata.

2.  **Extending the Dataset:** For each entry in the original dataset:
    * A new entry was created where the `original_triton_code` was placed in a new column named `original_triton_python_code`, and the `triton_code` was renamed to `optimised_triton_code`.
    * Another new entry was created where the `python_code` from the original entry was placed in the `original_triton_python_code` column, and paired with the same `triton_code` (now `optimised_triton_code`).
    * All other metadata columns (`entry_point`, `repo_name`, `module_name`, `synthetic`, `uuid`, `licenses`, `stars`, `sha`, `repo_link`, and potentially others) were retained for both new entries.

This extension strategy aims to provide training data where the model can learn to generate the same optimized Triton code from two different input representations: the original Triton code (potentially unoptimized or an intermediate form) and the higher-level Python/PyTorch code.



## Data Structure

Each entry in this dataset contains the following fields:

* **entry\_point**: The entry point of the code.
* **original\_triton\_python\_code**: Contains either the original Triton code or the Python code from the source dataset, representing the input to be optimized or translated.
* **optimised\_triton\_code**: The optimized Triton code generated by `torch.compile`.
* **repo\_name**: The name of the repository in the format `username/repository`.
* **module\_name**: The name of the PyTorch module.
* **synthetic**: A boolean indicating if the data is synthetic.
* **uuid**: A unique identifier for the entry.
* **licenses**: List of licenses associated with the repository.
* **stars**: Number of GitHub stars the repository has.
* **sha**: The commit SHA hash used for version reference.
* **repo\_link**: Direct link to the repository at the specific commit (GitHub URL).
* **stringlengths**: (Please inspect the data files for the exact content and meaning of this field, as it might contain summary statistics related to code lengths from the original dataset.)



## Usage Examples

You can load this dataset using the `datasets` library in Python:

```python
from datasets import load_dataset

extended_dataset = load_dataset("TEEN-D/Code_Opt_Triton")

# Example: Print the first 5 examples
print(extended_dataset['train'][:5])