Datasets:

---

SalutefaceAAC

/

structured-output-generation

Like 0

🎉 structured-output-generation v1.0 is live!

We're releasing structured-output-generation — 107 hand-authored natural-language prompts paired with ideal structured-format completions across JSON, YAML, CSV, Markdown tables, and XML.

Why this dataset?

Most instruction-following benchmarks focus on free-text responses. This dataset fills the gap for structured output tasks — a growing need as developers use LLMs to generate configs, API payloads, reports, and machine-readable data.

Highlights

• 107 examples across 5 formats and 15+ domains
• 3 difficulty levels: easy (flat, <7 fields), medium (nested, arrays), hard (RFC-compliant, computed fields, format-specific idioms)
• Realistic values — proper ISO 8601 timestamps, real ISBNs, plausible prices, standard HTTP codes
• Hard examples include: RFC 7807 error responses, JWT payloads, SOAP envelopes, Maven POMs, OpenAPI path definitions, Prometheus alerting rules, CI/CD pipeline runs
• JSONL + Parquet — native HF dataset viewer support

Format distribution

Quick start

from datasets import load_dataset

ds = load_dataset("SalutefaceAAC/structured-output-generation")
hard_json = ds["train"].filter(lambda x: x["format"] == "json" and x["difficulty"] == "hard")

Fine-tuning template

def to_prompt(example):
    return (
        f"Generate the following structured output:\n\n"
        f"{example['prompt']}\n\n"
        f"Format: {example['format'].upper()}\n\n"
        f"Output:\n{example['completion']}"
    )

Feedback, use cases, and contributions very welcome! 🙏