NL2SH-ALPACA (Alpaca-style)

This is a reformatted version of the NL2SH-ALFA dataset originally created by westenfelder/NL2SH-ALFA.

It has been converted to Alpaca-style format and prepared for instruction fine-tuning by Anshuman Jena, who acted as the converter and maintainer of this version.

{
  "instruction": "<natural language instruction>",
  "input": "",
  "output": "<bash command>"
}

Additionally, for the test split, the original bash2 (alternative command) and difficulty fields have been included in the output.

This dataset can be used to train instruction-following models for translating natural language instructions to shell commands.

---

Data Fields

The data fields are as follows:

• instruction: natural language description of the shell task the model should perform. Each instruction is unique.
• input: optional context or input for the task. In this dataset, this field is empty for all examples.
• output: the shell command corresponding to the instruction. For the test split, alternative commands (bash2) and difficulty are appended to the output.

---

Data Instances

An example of "train" looks as follows:

{
    'instruction': 'Compile C code and cache compiled output (to use `ccache` on all `gcc` invocations, see the note above)',
    'input': '',
    'output': 'ccache gcc path/to/file.c'
}

---

Data Splits

	train	test
NL2SH-ALPACA	40,639	300

---

How to load

from datasets import load_dataset
ds = load_dataset("abandonedmonk/NL2SH-ALPACA")
print(ds["train"][0])

---

Citation

If you use this dataset, please cite the original work:

@misc{westenfelder2025nl2sh,
title={NL2SH-ALFA},
author={westenfelder},
year={2025},
howpublished={\url{[https://huggingface.co/datasets/westenfelder/NL2SH-ALFA}}](https://huggingface.co/datasets/westenfelder/NL2SH-ALFA}})
}

This version of the dataset was converted to Alpaca-style format and maintained by Anshuman Jena.