stindardlogic/structured-output-sft-100k

Structured Output SFT (100K)

100,000 ShareGPT conversations demonstrating correct generation of structured data formats: JSON, YAML, CSV, XML, Markdown tables, JSON Schema, and OpenAPI fragments. Each example pairs a natural language specification with a valid, well-formed output.

Motivation

Structured output generation is among the most commercially critical LLM capabilities. Models fail in characteristic ways:

• Invalid JSON: unclosed brackets, trailing commas, unquoted keys, mismatched types
• Schema violations: required fields missing, wrong types, values outside enum constraints
• Nesting errors: objects where arrays expected, flat structure instead of hierarchy
• Format inconsistency: CSV with inconsistent quoting, XML without closing tags, YAML indentation errors
• Semantic errors: technically valid JSON that violates the intended schema

This dataset trains models to produce valid, semantically correct structured output from natural language specifications.

Dataset Description

100,000 conversations across 7 output formats and 20 task types:

Format Distribution

Format	Count	%
JSON	~36,400	36%
YAML	~18,200	18%
Markdown Tables	~18,200	18%
CSV	~13,600	14%
JSON Schema	~4,500	5%
Markdown (docs)	~4,500	5%
XML	~4,500	5%

Task Categories

Category	Examples
Data extraction (text to JSON)	4,546
Schema/config design	13,637
API design (request/response/error formats)	9,091
Infrastructure config (Docker, Kubernetes, CI)	13,637
Data generation (sample/test data)	9,091
Data formatting (tables, reports)	9,090
Data modeling (trees, state machines)	9,090
Technical documentation	4,545
Geospatial (GeoJSON)	4,545
ML results (confusion matrices)	4,545
Time-series data	4,545

Format Example

{
  "conversations": [
    {
      "from": "human",
      "value": "Extract the following information into JSON: 'John Smith, 34 years old, works as a software engineer at Acme Corp...'"
    },
    {
      "from": "gpt",
      "value": "```json\n{\n  \"name\": \"John Smith\",\n  \"age\": 34,\n  ...\n}\n```"
    }
  ],
  "metadata": {
    "format": "json",
    "topic": "user_profile_schema",
    "output_format": "json",
    "task": "extraction"
  },
  "id": "abc123"
}

Key Properties of Responses

1. Syntactic validity: Every JSON/YAML/CSV/XML output is syntactically valid and would parse without errors.

2. Schema adherence: When a schema is specified (required fields, type constraints, enum values, length limits), the output strictly follows it.

3. Semantic correctness: Beyond syntax, outputs are semantically meaningful -- GeoJSON coordinates are [lon, lat] not [lat, lon]; polygon rings are closed; state machine terminal states have no transitions.

4. Format-specific conventions: XML declarations, RSS version attributes, GeoJSON FeatureCollection wrapper, YAML --- document separator, CSV quoting -- each format's conventions are followed.

5. Explanatory notes: Where non-obvious choices are made (coordinate order, ring closure, terminal states), responses explain why.

6. Production-quality patterns: Kubernetes Deployments use readOnly, Docker Compose uses named volumes, GitHub Actions use pinned action versions (actions/checkout@v4).

Real-World Use Cases Covered

• REST API design: request bodies, response envelopes, error formats, pagination
• Infrastructure as Code: Kubernetes Deployments, Docker Compose, GitHub Actions CI/CD
• Data schemas: JSON Schema, OpenAPI component schemas, Pydantic-like definitions
• Data exchange: webhook payloads, GeoJSON features, RSS feeds
• Analytics: comparison tables, time-series, confusion matrices, cohort data
• Configuration: server config with nested structure, environment-specific settings
• Documentation: API endpoint docs with request/response examples

Use Cases for Training

• SFT fine-tuning for structured output reliability
• Training models for JSON mode and function calling
• Improving model performance on schema-following benchmarks
• Building AI data entry, extraction, and transformation pipelines
• Training code assistants that generate config files correctly
• Complementary to json-structured-output-dpo-3k (small DPO) -- this provides large-scale SFT

License

Apache 2.0