20 messy real-world LLM outputs paired with the JSON you're trying to extract. A go-to fixture set for testing tolerant JSON extractors (e.g. agentcast) — covers fenced/unfenced blocks, prose-wrapped JSON, refusals, malformed payloads, JSON5, JavaScript-mistaken-for-JSON, unicode, and edge primitives.
{
"id": "string",
"raw": "string", // exactly what the LLM produced
"expected": "object | array | null", // the… See the full description on the dataset page:
https://huggingface.co/datasets/mukunda1729/llm-output-extraction-cases.
