Subh775/Blockzie-NL2blocks

Blockzie NL2Blocks

Pairs of a plain-English request and the block program that answers it, for block-based visual programming.

About block coding and Blockzie

In block-based programming a program is built by snapping labelled blocks together on a canvas instead of typing text. Each block is one instruction (move 10 steps, wait 1 seconds, when green flag clicked), and blocks nest inside loops and conditionals. Scratch is the best-known editor of this kind.

Blockzie (https://blockzie.info) is an editor of the same family. Alongside the standard motion, looks, control and sensing blocks it adds two groups that ordinary Scratch does not have:

• Microcontroller blocks. A program can be uploaded to a board such as an ESP32 or an Arduino Uno, where it runs on the hardware and drives pins, sensors and displays; or it can run on the computer in realtime mode and talk to the board over a serial or Bluetooth link.
• Camera and AI extensions. Face detection, pose and hand tracking, object detection, QR scanning, speech to text, text to speech and translation.

A saved project is an .sb3 file: a zip holding project.json (the program and every sprite) and the costume and sound assets.

What this dataset is for

Models that write text code from a description are well covered by existing datasets; models that produce a block program are not. The gap matters because a block program is a tree of typed blocks with named inputs, not a sequence of tokens: a generator has to emit something the editor can load and place on the canvas.

Each row here gives a request and the finished program in the forms a model can be trained or evaluated on, so the same data supports:

• generating a program from a description, and placing it directly in an editor,
• retrieval examples for a prompted model,
• evaluating whether a generated program loads, runs and uses the blocks the request asked for.

The projects come from a taught robotics and AI course, so they lean towards hardware and camera work rather than arcade games.

Structure

data/train-*.parquet        106 rows
data/validation-*.parquet    10 rows
data/test-*.parquet          14 rows
sb3/bz-0001.sb3              one project file per row, openable in the editor
README.md                    this card

130 projects in total, holding 8,540 blocks: 48 camera and AI, 32 games, 32 hardware, 12 speech, 3 drawing, 1 physics, 2 other. 110 of the programs are distinct; the rest are the same project filed twice, marked so one filter removes them. Projects that are copies of each other stay in the same split.

Each row is one project.

Column	Type	Description
id	string	Stable row id, bz-0001.
query	string	The request, written the way a learner would ask for the program.
query_source	string	human if the request comes from the course material, generated if it was written from the project and reviewed.
description	string	The course's own description of the project.
objective	string	What the activity teaches, where recorded.
title	string	Activity name.
category	string	vision-ai, speech, hardware, game, drawing, animation, physics or data.
program_mode	string	realtime (runs on the computer) or upload (runs on the board).
board	string	Board the project targets, empty if none.
extensions	list[string]	Editor extensions the program uses.
device_extensions	list[string]	Board-side extensions.
project_json	string	The project's project.json: the program exactly as the editor saves it.
workspace_xml	string	The same program as workspace XML, one document per sprite, which is what an editor loads to place blocks on the canvas.
notation	string	The same program as indented one-block-per-line text (opcode NAME=value), compact enough to fit in a prompt.
sb3_path	string	Path of the project file in this repository.
sprites	list[string]	Sprite names.
sprite_count, block_count, script_count	int	Size of the program.
opcodes	list[string]	Block types used, for filtering by block.
duplicate_group	string	Shared by projects whose programs are identical.
is_canonical	bool	False on the extra copies of a duplicate group.
split	string	train, validation or test.

The three program forms

project_json is the ground truth, the exact file format of the editor. workspace_xml is the form a generator must emit to place blocks on a canvas. notation is the short form, useful as a model's output language when context is limited. All three are produced from the same project.json by one converter, so they always agree.

Usage

from datasets import load_dataset

ds = load_dataset("Subh775/Blockzie-NL2blocks", split="train")
row = ds[0]
print(row["query"])
print(row["notation"])

Coverage and limits

• Small: 130 projects, suited to retrieval, evaluation and light fine-tuning rather than training a model from scratch.
• Subject balance follows the course: many camera-AI and hardware projects, few drawing projects.
• query is currently derived from each activity's own title and description, and marked derived in query_source. Requests written or checked by hand will replace them, and will be marked human.
• Descriptions were written for teaching, so they say what a project does and not always how it is built. Six are shorter than one line.
• Assets are not included: every sprite carries one stock costume and no sounds, so a project opens and runs but does not look like the original.

Provenance

Built from the project files and activity sheets of an educational block-coding curriculum. Programs are unchanged; metadata was normalised and personal and internal identifiers removed.

Licence

MIT.