Datasets:

LaelaZ
/

synthkit-demo

license: mit
language:
  - en
pretty_name: 'SynthKit Demo: Synthetic Coding-Instruction Prompts'
size_categories:
  - n<1K
task_categories:
  - text-generation
tags:
  - synthetic
  - synthetic-data
  - code
  - coding
  - instructions
  - instruction-tuning
  - prompts
  - data-quality
  - benchmark
  - contamination
configs:
  - config_name: default
    data_files:
      - split: train
        path: train.jsonl
      - split: benchmark
        path: benchmark.jsonl

SynthKit Demo: Synthetic Coding-Instruction Prompts

Anyone can generate synthetic data. The hard part is knowing whether it's any good, or whether your eval set has leaked into your training set without you noticing. This small dataset is the demo for SynthKit, a tool that generates data and then grades it before you train on it. Try the grader in your browser: 🤗 huggingface.co/spaces/LaelaZ/synthkit.

The point isn't the size. It's the setup. The benchmark split overlaps the training split on purpose, by exactly five records, so the dataset doubles as a worked example of catching train/eval contamination instead of just talking about it.

What's in it

Split	Records	What it is
`train`	200	synthetic coding-instruction prompts
`benchmark`	40	held-out eval prompts, 5 of which also sit in `train` (planted)

Schema (both splits): one record looks like this.

{"prompt": "How would you reverse a string in TypeScript? Walk through your reasoning.", "domain": "coding"}

prompt (string): a coding instruction or question.
domain (string): task domain (coding throughout this demo).

The prompts are templated variants over about ten canonical algorithm tasks (binary search, cycle detection, moving averages, prime checking, list flattening, and so on) crossed with eight languages (C++, Rust, Python, JavaScript, TypeScript, Go, Java, Ruby) and a few framings: explain, refactor, debug, reason it through.

What the grader says about it

SynthKit grades what it makes, so this demo comes with its own scorecard, and you can reproduce it from a clean checkout:

Axis	Score	What it measures
Overall	B (89.8 / 100)	headline grade
Validity	100	200/200 records well-formed
Uniqueness	92	185/200 unique (0 exact, 15 near-duplicate)
Diversity	70	leans on pairwise self-similarity, since distinct-n shrinks as a set grows
Contamination	98	flags exactly the 5 records that overlap the benchmark, no more, no fewer

Usage

Each split is selectable. Load what you need:

from datasets import load_dataset

ds = load_dataset("LaelaZ/synthkit-demo")
train     = ds["train"]       # 200 prompts
benchmark = ds["benchmark"]   # 40 prompts, 5 planted in train

# Reproduce the contamination signal: which prompts sit in both splits?
overlap = set(r["prompt"] for r in train) & set(r["prompt"] for r in benchmark)
print(len(overlap), "contaminated prompts")  # -> 5

How it was generated

SynthKit's template provider built it: deterministic template substitution over a small task / language / framing grid, then a near-duplicate filter. It's a pure function of the seed templates (examples/*.json in the repo), runs on the Python standard library alone, and contains no real, scraped, or personal data.

What it's good for, and what it isn't

Use it as a tiny, reproducible fixture for demonstrating data-quality grading, validity, uniqueness, diversity, and contamination, or as a quick instruction-prompt sample for a tutorial.

Don't use it to train a model. It's small (240 prompts), single-domain (coding), template-generated so its lexical diversity is bounded by design, and the prompts have no reference answers attached. It's a demonstration of grading, not a training corpus.

License & citation

@misc{zorana_synthkit_demo_2026,
  author = {Laela Zorana},
  title  = {SynthKit Demo: Synthetic Coding-Instruction Prompts},
  year   = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/LaelaZ/synthkit-demo}}
}

Links: SynthKit on GitHub · live grader Space · related: LaelaZ/synthetic-ecommerce