Emulated-Inc/json-schema-instances-training-pool

JSON schema and instance training pool

Real JSON Schemas from the public collections named below, read at the pinned revisions given there, each paired where possible with documents that satisfy it, laid out twice. Train on either layer or on both.

pool.jsonl

Every source rewritten into one shape, 20004 rows, one JSON object per line, with these fields.

Field	What it holds
id	a row identifier unique within this file
prompt	the request a model would answer: one line, the schema, one line
schema	the JSON Schema the document has to satisfy
documents	a list of JSON documents that satisfy the schema, empty when none was built
document_origin	rule-generated when the row carries documents, null otherwise
row_shape	schema_documents or schema, which of the two the row is
source	the name of the source directory the row came from
source_repo, source_revision	the collection and the revision it was read at
source_subset, source_id	where the row sits in that collection
provenance_class	how the schema came to exist
licence	the licence of the material the row came from
root_kind	the kind of value the schema's root admits: object, array, a scalar type name, or unknown when the declared keywords do not say

Which rows carry what. 13712 rows carry at least one document and 6292 rows carry the schema alone. 23091 documents in all.

Every schema in this file is accepted by the validator as a schema of the draft it declares, and every document validates against the schema of its own row. Rows are deduplicated on the schema with its human annotations stripped, keeping the first source that carries it in the order of the sections below.

Where the documents come from. No public collection publishes an instance beside its schema, so the documents here were written by a rule-based generator that reads a schema and emits a value satisfying it, and each one was checked against its schema before it was kept. They are declared rule-generated for that reason. They are legal instances and they are not natural data: a generator fills a string with a run of letters and a number with the smallest value the bounds allow. Up to two documents are written per schema, one with the optional keys left out and one with them filled, where the two differ.

sources/

The same material untouched, 25165 records, one directory per source, holding the files at the paths, in the format and with the fields its own repository publishes. Nothing here was renamed, reshaped, reordered, validated or deduplicated, and no document was added. Use this layer if you want a field the rewritten one drops, such as the repository and star count of a scraped schema, or if you would rather decide for yourself which schemas to keep.

The sources

sources/json_schema_corpus

Real JSON Schemas collected from public source repositories, from the Kubernetes API, from the Snowplow event registry, from the community schema catalogue, from the Washington Post content specification and from the FHIR health-record specification, graded into difficulty tiers, plus the release's own hand-written set.

• Read from epfl-dlab/JSONSchemaBench at 5bd0f4640badc6f3f02df796421d21cb0ca0b141.
• Licence: mit.
• How it came to exist: collected.
• 6605 records kept of 8154 read, 6317 rows in the rewritten layer.
• The difficulty tiers are the release's own and are kept as its own subset field.

sources/function_call_schemas

Function parameter schemas lifted out of a synthetic function-calling corpus and republished as standalone schemas.

• Read from epfl-dlab/JSONSchemaBench at 5bd0f4640badc6f3f02df796421d21cb0ca0b141.
• Licence: mit.
• How it came to exist: model-generated.
• 989 records kept of 1707 read, 544 rows in the rewritten layer.
• These schemas were invented by a model rather than harvested, which is why they are a source of their own. They are narrower and more uniform than real ones.

sources/draft_conformance_schemas

Small schemas written by a generator to exercise one keyword combination each, from the same release.

• Read from epfl-dlab/JSONSchemaBench at 5bd0f4640badc6f3f02df796421d21cb0ca0b141.
• Licence: mit.
• How it came to exist: rule-generated.
• 413 records kept of 413 read, 413 rows in the rewritten layer.
• These are puzzles rather than descriptions of data: most of them describe a scalar rather than an object, and many are contradictory on purpose.

sources/github_schemas

JSON Schema files scraped from public repositories through a code-search API, every revision fetched and validated, split so that the schemas of one organisation stay together.

• Read from dataunitylab/json-schema at fc9d99ae0ad504e85b9c03e4db0929c9b97d283b.
• Licence: unknown, per-row permissive.
• How it came to exist: collected.
• 16275 records kept of 16367 read, 11993 rows in the rewritten layer.
• The repository-level licence tag is literally unknown, which is why this source is flagged. Every row nevertheless names the licence of the repository it came from and every one of those is permissive, the collection having been filtered to a permissive allowlist, so the rows carry their own.

sources/schema_catalogue

The community catalogue of JSON Schemas for configuration and data formats, the set that editors and language servers resolve against, written and maintained by the authors of the formats and by the catalogue's contributors.

• Read from SchemaStore/schemastore at 04787bb09aec3add05cf75889c5d1f4cba9625f1.
• Licence: apache-2.0.
• How it came to exist: collected.
• 883 records kept of 977 read, 737 rows in the rewritten layer.
• One file per format, so the schemas are large and heavily documented, and many of them reference sibling files of the catalogue. Its raw layer is one schemas.jsonl.gz holding the path and the content of every file, rather than the thousand files themselves.

Licence

The pool carries apache-2.0, the most restrictive term its sources compose to. Each row and each source directory names its own, so a subset under a single term can be selected. The scraped collection carries no repository-level term and every one of its rows names the permissive licence of the repository the file came from.