arunsrajan/godot-docs-dataset

Godot Engine Docs Q&A

24,631 instruction/answer pairs generated from the official Godot Engine documentation (4.7 branch), for fine-tuning LLMs such as Qwen, Gemma and Llama with Unsloth.

Generated with godot-docs-to-csv.

Columns

Column	Description
instruction	Question built from a template, e.g. "Explain the add_child method of the Node class in Godot."
input	Always empty (kept for Alpaca compatibility)
output	The documentation text in Markdown, with GDScript / C# / GLSL samples in fenced code blocks
category	Docs section the row comes from, e.g. classes, tutorials/2d
source	Page and anchor in the docs, e.g. classes/class_node.html#class-node-method-add-child

Contents

Category	Rows	What a row covers
classes	20,811	A class overview, property, method, signal, enum, constant, constructor, operator or theme property, or a list of a class's properties/methods
tutorials	3,103	One section of a tutorial or manual page
engine_details	416	One section of the engine internals docs
getting_started	153	One section of the getting started guide
about, community, root pages	148	One section of the remaining pages

Undocumented class members are left out. Answers longer than 4,000 characters are split into (part i/n) rows without breaking code blocks.

Limitations: questions come from fixed templates, and answers are the documentation text itself. Class reference rows make up about 85% of the data, so downsample them if you want more tutorial-style answers. The docs are from a development branch, so a few APIs may change before release.

Usage

Pass keep_default_na=False, otherwise the empty input column is read as None.

from datasets import load_dataset

dataset = load_dataset("arunsrajan/godot-docs-dataset", split="train", keep_default_na=False)

def to_chat(batch):
    texts = []
    for question, answer in zip(batch["instruction"], batch["output"]):
        messages = [{"role": "user", "content": question},
                    {"role": "assistant", "content": answer}]
        texts.append(tokenizer.apply_chat_template(messages, tokenize=False))
    return {"text": texts}

dataset = dataset.map(to_chat, batched=True)  # then train with SFTTrainer on the "text" field

License and attribution

The content is derived from the Godot Engine documentation by Juan Linietsky, Ariel Manzur and the Godot community, licensed under CC BY 3.0. This dataset is released under the same license. The text has been reformatted (HTML converted to Markdown, split into question/answer pairs).