respinosamena/Helios-Nexus-JSON-Data

---
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

```python
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