README.md

---
license: mit
task_categories:
  - text-generation
language:
  - en
tags:
  - spelling
  - grammar
  - typo-correction
  - phonetic
  - evaluation
size_categories:
  - n<1K
pretty_name: Open Spell Evaluation Set
gated: manual
extra_gated_prompt: >
  Access to the Open Spell evaluation dataset is granted on request.
  Please tell us how you intend to use it so we can route the approval correctly.
extra_gated_fields:
  Company: text
  Use case: text
  Email: text
extra_gated_button_content: Request access
---

# Open Spell — Evaluation Dataset

A small, hand-curated set of misspelled / corrected pairs used to validate the **Open Spell**
correction kit across four categories: plain spell, grammar (segmentation), phonetic, and domain.

> Not a training set. SymSpell is non-parametric and does not need one. This dataset is for
> regression testing and reporting top-1 accuracy.

## Schema

Each line of `samples.jsonl` is one example:

```json
{
  "id": "spell-001",
  "category": "spell",
  "original": "I lik aples",
  "corrected": "I like apples",
  "domain_terms": []
}
```

| Field | Type | Description |
|-------|------|-------------|
| `id` | string | `{category}-{index}` |
| `category` | enum | `spell`, `grammar`, `phonetic`, or `domain` |
| `original` | string | Input text with errors |
| `corrected` | string | Expected output |
| `domain_terms` | string[] | Vocabulary to load before correcting (`domain` category only) |

## Loading

```python
import json
from pathlib import Path

samples = [json.loads(l) for l in Path("samples.jsonl").read_text().splitlines()]
```

Or with `datasets`:

```python
from datasets import load_dataset
ds = load_dataset("json", data_files="samples.jsonl")
```

## Categories

| Category | Count | What it tests |
|----------|-------|---------------|
| `spell` | 15 | Single-word typos, capitalization, trailing punctuation |
| `grammar` | 15 | Missing spaces, extra spaces, compound segmentation |
| `phonetic` | 10 | Sounds-like substitutions (`fone` → `phone`) |
| `domain` | 10 | Words only correctable with injected custom vocabulary |

## License

MIT. Examples written from scratch — no scraped or copyrighted material.