LaelaZ/synthetic-instruction-prompts

Synthetic Instruction Prompts (8 domains)

Most synthetic prompt sets are a black box. You get a pile of prompts and no idea whether they're actually varied or just the same three sentences wearing different nouns. This one is graded, and the grade is on the card.

It's 2,829 instruction-style prompts across eight domains, generated with SynthKit and then scored by the same tool. The prompts carry no answers. Think of them as seed prompts: you feed them to a model to bootstrap instruction-tuning data, or you use them as a quick eval set. Generation was offline, free, and seeded, so the whole dataset rebuilds byte-for-byte from the script included here.

Domains

Eight domains, weighted toward code because that's where the deepest template space was.

Domain	Prompts	Example
coding	859	"What's the most efficient way to implement an LRU cache in Rust, and why?"
data-science	480	"How do I tune hyperparameters with cross-validation in scikit-learn? Explain each step."
devops	320	"What's a best practice for a zero-downtime rolling update with Kubernetes?"
debugging	280	"I'm getting this error in Python: "IndexError: list index out of range". What's the likely cause and fix?"
sql	280	"Write a PostgreSQL query to compute a 7-day rolling average of sales."
technical-writing	260	"Outline a runbook that explains a backup-and-restore procedure."
customer-support	200	"Draft a reply to a frustrated customer who says: "I was charged twice this month""
math-word-problems	150	"Solve this for a high-school student and show your steps: two pipes fill a tank in 6 and 9 hours; how long together?"
Total	2,829

Splits

Split	Records
train	2,405
test	424

The test split is a deterministic 15% hold-out. The two splits share no prompt: the exact-prompt overlap between them is zero. They do share template scaffolding, though, since they come from the same generator, so treat the hold-out as a hold-out, not as an independent benchmark.

Schema

{"prompt": "Write a Go function that detects a cycle in a linked list.", "domain": "coding"}

• prompt (string): an instruction-style prompt.
• domain (string): one of the eight domains above.

What the grader says about it

SynthKit scored the full set, and the report is in quality_report.json:

Axis	Score	Read it like this
Overall	B (86.2 / 100)	headline grade
Validity	100	2,829/2,829 well-formed
Uniqueness	90.9	2,572 unique, 0 exact and 257 near-duplicate
Diversity	66.7	distinct-1 0.015 · distinct-2 0.043 · self-similarity 0.004
Contamination	n/a	not measured, because there's no external benchmark to check against

Here's what those numbers actually mean, because the grade is honest about a real limitation. The prompts are template-generated, so a lot of them share scaffolding and differ only by a slot: the language, the library, the task. That's why the near-duplicate count isn't zero and the diversity score sits in the sixties. The distinct-n figures also fall on their own as any corpus grows, which is why diversity leans on the self-similarity number instead (0.004 is very low, which is the good direction). So: a diverse-enough seed bank, not a high-entropy corpus. That's the honest framing.

Usage

from datasets import load_dataset

ds = load_dataset("LaelaZ/synthetic-instruction-prompts")
train, test = ds["train"], ds["test"]

# Pull one domain
sql = train.filter(lambda r: r["domain"] == "sql")

# The usual move: send each prompt through your own model to get the answers
for r in train.select(range(3)):
    print(r["domain"], "|", r["prompt"])

How it was generated

Eight SynthKit seed specs (templates with {slot} placeholders plus filler lists) were sampled with provider=none, which produces the prompt only and costs nothing. Then the prompts were deduplicated, pooled, and split. It runs on the Python standard library alone and is a pure function of the seeds plus the RNG seed (17). The generator comes with the dataset as generate.py, so pip install synthkit && python generate.py rebuilds every file here, byte-for-byte. Edit the SPECS dict to add your own domains.

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

Use it as seed prompts for synthesizing instruction-tuning data (run them through your own model to get responses), as a lightweight multi-domain eval prompt set, or as a worked example of grading synthetic data.

Don't train a model on it directly, because there are no responses. It's also template-generated, so the lexical diversity is bounded, and it's English-only. Use it to seed and to test, not to measure a model's capability.

License & citation

MIT © 2026 Laela Zorana.

@misc{zorana_synth_instruction_prompts_2026,
  author = {Laela Zorana},
  title  = {Synthetic Instruction Prompts (8 domains)},
  year   = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/LaelaZ/synthetic-instruction-prompts}}
}

Links: SynthKit on GitHub · live grader Space · siblings: synthkit-demo · synthetic-ecommerce