SalutefaceAAC/structured-output-generation

Structured Output Generation

A hand-authored synthetic dataset of 107 natural-language prompts paired with ideal structured-format completions, spanning JSON, YAML, CSV, Markdown tables, and XML. Designed for fine-tuning and evaluating instruction-following models on structured output tasks.

Dataset at a Glance

Property	Value
Total examples	107
Formats	JSON (36), YAML (20), Markdown table (19), CSV (18), XML (14)
Difficulty levels	Easy (26), Medium (46), Hard (35)
Domains	software, finance, healthcare, AI, education, HR, ecommerce, logistics, IoT, media, sports, geography, food, legal, science
Train / Val / Test	85 / 10 / 12
License	Apache 2.0
Generation method	Hand-authored

Schema

Each example has 8 fields:

Field	Type	Description
id	string	MD5 hash of prompt+completion (12 chars)
prompt	string	Natural-language instruction requesting structured output
format	string	Target format: json, yaml, csv, markdown_table, xml
schema_description	string	One-line description of the expected output structure
completion	string	The ideal structured output
domain	string	Subject domain (software, finance, healthcare, …)
difficulty	string	easy, medium, or hard
num_fields	int	Approximate number of top-level fields / columns

Sample Records

JSON — Medium difficulty (finance)

Prompt:

Produce a JSON object for a bank transaction record. Fields: transaction_id, account_from, account_to, amount, currency, type (debit/credit/transfer), and timestamp.

Completion:

{
  "transaction_id": "TXN-20241101-998812",
  "account_from": "ACC-001-882234",
  "account_to": "ACC-002-117765",
  "amount": 2500.00,
  "currency": "USD",
  "type": "transfer",
  "timestamp": "2024-11-01T09:15:33Z"
}

---

YAML — Hard difficulty (software)

Prompt:

Write a YAML config for a Kubernetes Deployment of an nginx container with 3 replicas, a CPU request of 100m and limit of 500m, memory request of 128Mi and limit of 256Mi, and a readiness probe on port 80 path /healthz.

Completion:

apiVersion: apps/v1
kind: Deployment
metadata:
  name: nginx-deployment
spec:
  replicas: 3
  selector:
    matchLabels:
      app: nginx
  template:
    metadata:
      labels:
        app: nginx
    spec:
      containers:
        - name: nginx
          image: nginx:1.27
          resources:
            requests:
              cpu: "100m"
              memory: "128Mi"
            limits:
              cpu: "500m"
              memory: "256Mi"
          readinessProbe:
            httpGet:
              path: /healthz
              port: 80
            initialDelaySeconds: 5
            periodSeconds: 10

---

CSV — Medium difficulty (ecommerce)

Prompt:

Create a CSV of 5 e-commerce orders. Columns: order_id, customer_email, order_date, item_count, subtotal_usd, shipping_usd, tax_usd, total_usd, status.

Completion:

order_id,customer_email,order_date,item_count,subtotal_usd,shipping_usd,tax_usd,total_usd,status
ORD-10001,alice.chen@email.com,2024-10-29,2,89.98,5.99,7.65,103.62,delivered
ORD-10002,bob.wilson@email.com,2024-10-30,1,24.99,0.00,2.12,27.11,shipped
ORD-10003,diana.osei@email.com,2024-10-31,4,215.50,12.99,18.32,246.81,processing
ORD-10004,ethan.ko@email.com,2024-11-01,1,149.99,9.99,12.75,172.73,pending
ORD-10005,fiona.ray@email.com,2024-11-01,3,67.47,5.99,5.73,79.19,confirmed

---

Markdown Table — Hard difficulty (ai)

Prompt:

Create a Markdown results table for a benchmark comparing three LLM inference backends (vLLM, TGI, LiteLLM) on: throughput (tokens/sec), latency P50 (ms), latency P99 (ms), GPU memory (GB), and max concurrent requests.

Completion:

| Backend | Throughput (tok/s) | Latency P50 (ms) | Latency P99 (ms) | GPU Memory (GB) | Max Concurrent |
|---|---|---|---|---|---|
| vLLM | 3,840 | 128 | 410 | 18.2 | 512 |
| TGI | 2,970 | 165 | 520 | 16.8 | 256 |
| LiteLLM | 1,820 | 243 | 780 | 14.1 | 128 |

---

XML — Medium difficulty (iot)

Prompt:

Write an XML document for an IoT device heartbeat message. Include: device_id, device_type, firmware_version, timestamp, battery_level_pct, signal_strength_dbm, sensors (list of sensor elements with id and status), and location (lat, lon).

Completion:

<?xml version="1.0" encoding="UTF-8"?>
<heartbeat>
  <device_id>DEV-IOT-8812</device_id>
  <device_type>environmental_monitor</device_type>
  <firmware_version>3.2.1</firmware_version>
  <timestamp>2024-11-01T14:00:00Z</timestamp>
  <battery_level_pct>74</battery_level_pct>
  <signal_strength_dbm>-62</signal_strength_dbm>
  <sensors>
    <sensor id="temp-01" status="ok"/>
    <sensor id="humidity-01" status="ok"/>
    <sensor id="co2-01" status="degraded"/>
  </sensors>
  <location>
    <lat>51.5074</lat>
    <lon>-0.1278</lon>
  </location>
</heartbeat>

Usage

from datasets import load_dataset

ds = load_dataset("SalutefaceAAC/structured-output-generation")
train = ds["train"]

# Filter by format
json_examples = train.filter(lambda x: x["format"] == "json")

# Filter by difficulty
hard_examples = train.filter(lambda x: x["difficulty"] == "hard")

Fine-tuning prompt 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']}"
    )

Generation Methodology

All examples were hand-authored to ensure:

1. Prompt clarity — each prompt unambiguously specifies the desired output format and fields
2. Completion correctness — all JSON/YAML/XML is syntactically valid; CSV headers match column descriptions; Markdown tables render correctly
3. Domain diversity — 15+ domains including software, finance, healthcare, AI/ML, education, ecommerce, IoT, sports, legal, and science
4. Realistic values — data uses plausible names, codes, prices, timestamps, and identifiers rather than placeholder values
5. Format coverage — hard examples include nested structures, arrays of objects, references between fields, and format-specific idioms (SOAP envelopes, Atom feeds, Maven POMs, OpenAPI paths)

Difficulty Levels

• Easy: Simple flat structures, 5-7 fields, no nesting required
• Medium: Moderate nesting (1-2 levels), some arrays, domain-specific conventions
• Hard: Deep nesting, arrays of objects, format-specific idioms, cross-field consistency (e.g. computed totals, RFC-compliant structures)

Limitations

• English-only prompts and completions
• ~107 examples — sufficient for evaluation and few-shot prompting; augmentation recommended for full SFT
• Software domain is intentionally over-represented (~39%) reflecting real-world structured output demand
• XML examples do not cover DTD/XSD validation patterns
• Does not include adversarial or out-of-distribution prompts

Intended Uses

• Supervised fine-tuning (SFT) of instruction-following models for structured output generation
• Evaluation of model ability to produce format-valid and semantically correct structured data
• Few-shot prompting baselines for structured generation tasks
• Curriculum data for difficulty-staged training

License

Apache 2.0 — free to use, share, and modify with attribution.

Citation

@dataset{salutefaceaac2024structured,
  title  = {Structured Output Generation},
  author = {SalutefaceAAC},
  year   = {2024},
  url    = {https://huggingface.co/datasets/SalutefaceAAC/structured-output-generation}
}