Datasets:

fromziro
/

Rephrased2JSON-Preview

Rephrased2JSON-Preview is a synthetically generated, cross-domain dataset designed to train large language models on structured information. The dataset transforms unstructured natural language into semantic JSON representations capturing key facts, named entities, relationships, abstract concepts, and quantitative metrics.

This preview release contains ~20,000 high-density extraction records generated locally using Liquid AI's LFM2.5-1.2B-Instruct.

Schema structure

{
  "doc_id": "dclm_14028",
  "text": "... [Source passage truncated up to 2,048 tokens] ...",
  "text_len": 1102,
  "text_from": "HuggingFaceTB/dclm-edu",
  "results": [
    {
      "generated_output": "... [generated output] ...",
      "len": 384,
      "valid": true,
      "grammar_constrained": true,
      "type_profile": {
        "numbers": 4,
        "strings": 28,
        "arrays": 4,
        "objects": 6,
        "booleans": 0
      }
    }
  ]
}

Generated Output Example

{
  "location": "South Jamaica housing projects, Queens",
  "subject": "Milford Graves",
  "age_context": "young man",
  "residence_address": "110th Avenue",
  "property_features": [
    "ornate mosaic",
    "reflective metal",
    "discarded marble",
    "lush garden",
    "citrus trees",
    "herbs",
    "exotic plants"
  ],
  "notable_achievements": [
    "1967 Down Beat magazine talent poll honor",
    "Offers from Miles Davis and Miriam Makeba",
    "Created computer programs to analyze heart rhythms",
    "Developed biofeedback for heart rhythm correction"
  ],
  "medical_impact": [
    "Helps detect heart problems",
    "Potential to cure heart issues",
    "Uses musical composition to influence heart sound",
    "Applied to patients like Dennis Thomas"
  ],
  "collaborations": [
    "Harvard Medical School",
    "North Shore University Hospital",
    "John Simon Guggenheim Memorial Foundation"
  ],
  "expert_opinions": [
    "Dr. Baruch Krauss: Recognizes Graves as a Renaissance man bridging music and medicine",
    "Dr. Ram Jadonath: Supports the idea of musical rhythm in cardiology",
    "Dr. Zorn: Compares Graves to a 20th-century shaman"
  ],
  "techniques_used": [
    "Biofeedback",
    "Algorhythmic formula",
    "Composing from heartbeat melodies",
    "Acupuncture point stimulation"
  ],
  "quantitative_outcomes": [
    "10 beats per minute increase in patient heart rate",
    "Improved breathing and relaxation reported by patient"
  ],
  "additional_context": "Graves lives in a basement with a garden and uses music as both art and medicine."
}

How to load

from datasets import load_dataset

# Load the full dataset
dataset = load_dataset("fromziro/Rephrased2JSON-Preview", split="train")

# or stream without downloading
# dataset = load_dataset("fromziro/Rephrased2JSON-Preview", split="train", streaming=True)

print(dataset[0])

Notes

When training on generated outputs, do not train on the "text" field. Train on the "results[].generated_output" field.
Filter by results[].valid == true if you only want syntactically valid JSON.
For extraction tasks, use "text" as the input prompt and "results[].generated_output" as the target output.