TGPRO32/Paragon-coding

NOTICE

This was done by me, someone with a learning Disability. So please do bare with me when updating this with more working data.

Multi-Language Programming Code Dataset

A curated dataset of original, non-scraped code examples across 7 programming environments: Python, JavaScript, Node.js, Java, C, C++, and Rust.

The dataset ships in two parts that can be used separately or combined:

Note on Node.js: Node.js is a JavaScript runtime, not a separate language. It's included as its own split because it exposes different APIs (filesystem, Buffer, process, http, CommonJS modules) than browser-context JavaScript — which is usually what people actually mean by "Node.js code."

Dataset Structure

Both files share the same schema:

code_dataset (105 rows) — Core Concepts

15 categories × 7 languages, one example each: Hello World · Variables and Data Types · Control Flow · Loops · Functions · Arrays and Lists · Dictionaries and Maps · Classes and OOP · Error Handling · File I/O · String Manipulation · Recursion · Sorting Algorithm · Async and Concurrency · Data Structures

code_dataset_large (2,255 rows) — Deep Coverage on 3 Categories

Generated by varying real parameters — algorithm choice, data type, sample values, operation sequences, and identifier names — not by duplicating templates with find-and-replace. Breakdown:

Distribution is balanced across languages (~320–326 rows each) and skews intermediate (1,727) over beginner (528), reflecting the algorithmic focus of this batch. Exact-duplicate rows were checked and removed (~3% collision rate from small-integer arrays landing on the same random sample).

Files

• —code_dataset.jsonl / code_dataset_large.jsonl — one JSON object per line (recommended for datasets.load_dataset("json", ...))
• —code_dataset.csv / code_dataset_large.csv — same data, spreadsheet-friendly
• —generate_dataset.py — generator for the 105-row core set (add more languages/categories by adding add(...) calls)
• —generate_batch.py — generator for the 2,255-row deep-coverage set (add more categories/algorithms by extending the template dicts)

Provenance & License

All code was written from scratch (hand-authored for the core set; programmatically templated with varied real parameters for the large set) — nothing was scraped from GitHub or any other source, so there are no third-party license conflicts. Released under MIT — free to use, modify, and redistribute, including for model training.

Known Limitations

• —The 2,255-row set currently covers only 3 categories in depth (sorting, data structures, strings). Categories like "Hello World" or "Variables" don't have enough genuine variation to scale the same way — padding them would mean shallow repetition rather than useful diversity.
• —For large-scale pretraining, pair this with an established corpus like The Stack or CodeSearchNet.
• —Snippets favor clarity/idiom over production hardening (minimal input validation) — they teach the pattern, not production-ready code.

Example Rows

Core set:

{
  "id": 1,
  "language": "Python",
  "category": "Hello World",
  "difficulty": "beginner",
  "task_description": "Print 'Hello, World!' to the console.",
  "code": "print(\"Hello, World!\")",
  "explanation": "Python's print() function writes text to standard output."
}

Large set:

{
  "id": 11,
  "language": "Python",
  "category": "Sorting Algorithm",
  "difficulty": "intermediate",
  "task_description": "Sort a 12-element array of ints in ascending order using bubble sort.",
  "code": "def bubble_sort(entries):\n    n = len(entries)\n    ...",
  "explanation": "Bubble sort on int data, variable named 'entries', 12 elements."
}

Loading

Hugging Face `datasets`:

from datasets import load_dataset
core = load_dataset("json", data_files="code_dataset.jsonl")
large = load_dataset("json", data_files="code_dataset_large.jsonl")

Pandas / Kaggle:

import pandas as pd
core = pd.read_csv("code_dataset.csv")
large = pd.read_csv("code_dataset_large.csv")
combined = pd.concat([core, large], ignore_index=True)

Suggested Uses

• —Fine-tuning a code-explanation or code-generation model
• —Few-shot prompting examples for a coding assistant
• —Cross-language idiom comparison (e.g., "how does error handling differ between Python and Rust?")
• —Algorithm-variant training data (many sorting/data-structure/string examples with controlled, labeled variation)
• —A regression-test seed set for code-generation model evals

Roadmap

The large set can be extended the same way to more categories (recursion, OOP, error handling, file I/O, async) by adding template functions to generate_batch.py — happy to keep scaling this up on request.