README.md

metadata

language:
  - en
license: apache-2.0
tags:
  - information-extraction
  - json
  - rag
  - structured-data
  - synthetic
  - legacy-database-modernization
task_categories:
  - text-generation
  - feature-extraction
size_categories:
  - 1B<n<10B
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*.parquet

Helios Nano JSON Data

Large-scale synthetic dataset for training small language models (SLMs) on structured information extraction — converting unstructured text into JSON.

Purpose

Designed for fine-tuning a 400M-parameter extraction engine that:

• Reads unstructured business documents (invoices, medical records, contracts, etc.)
• Follows a provided JSON schema
• Outputs clean, structured JSON

Ideal for legacy database modernization and RAG pipelines.

Dataset Structure

Each row contains:

Column	Type	Description
industry	string	Source industry (finance, healthcare, hr, legal, …)
doc_type	string	Document type (invoice, prescription, contract, …)
schema_json	string	JSON schema the model should extract
raw_text	string	Unstructured source document
extracted_json	string	Gold-standard structured extraction

Coverage

16 industries, 41 document types, including:

• Finance: invoices, receipts, payroll, wire transfers, tax summaries, bank transactions
• Healthcare: patient records, prescriptions, lab results, referrals
• HR: employee records, job postings, performance reviews
• Legal: contract summaries
• Real Estate: property listings, lease agreements
• Logistics: shipping notices, purchase orders, inventory, customs declarations
• Retail: orders, returns
• Insurance: claims
• Education: enrollment, scholarships
• Manufacturing: quality inspections, maintenance logs
• Government: business licenses, building permits
• And more…

Format Diversity

Text fields use randomized formatting for dates (Sept 29 / 09-29-2024 / 2024-09-29), currency ($1,234.56 / USD 1234.56), phone numbers, IDs, and document layout (formal headers vs. narrative prose vs. email style).

Stats

• Shards: 26
• Disk size: 12.2 GB (Snappy-compressed Parquet)
• Target: 10B tokens (BPE, vocab 32768)

Usage

from datasets import load_dataset

ds = load_dataset("respinosamena/Helios-Nano-JSON-Data", split="train")
print(ds[0])

Training Prompt Format

<|schema|>{schema_json}<|end_turn|>
<|document|>{raw_text}<|end_turn|>
<|extract|>{extracted_json}<|end_turn|>

License

Apache 2.0