IB-Emper/hinglish-translit-sarvam

Hinglish -> Devanagari Sloppy Transliteration Pairs (Synthetic, Sarvam-30b)

Synthetic dataset of realistic, sloppy romanized-Hindi ("Hinglish") spellings mapped to their correct Devanagari word or short phrase. Built to train a lightweight input-method engine (IME) that accepts fuzzy typed roman input and outputs correct Hindi script, in the style of a Chinese Pinyin smart-input keyboard.

How this was built

1. Ground truth is never model-generated. Seed words and short phrases (2-4 word n-grams) were pulled from a real Hindi corpus (Leipzig Corpora Collection, hin_news_2011_1M), frequency-ranked -- never invented by an LLM.
2. Each seed item was sent to Sarvam-30b with one of four rotating "typing persona" prompts (fast/careless thumbs, phonetic confusion, dropped-schwa casual texting, inconsistent casual), asking it to produce several plausible SLOPPY romanizations a real typist would produce for that exact word/phrase -- never asked to invent or correct the ground truth itself.
3. Every returned variant was validated: must be lowercase roman a-z (+ single spaces for phrases), must echo back a real seed item, invalid or hallucinated entries were dropped.
4. Exact duplicate (roman, devanagari) pairs across personas/batches were collapsed into one row with a count field.

Data Fields

• roman (string): the sloppy romanized input, e.g. "namste"
• devanagari (string): the correct Hindi word or phrase, e.g. "नमस्ते"
• kind (string): "word" or "phrase" (2-4 words)
• persona (string): which typing-persona prompt first produced this pair (fast_thumbs, phonetic_confusion, dropped_schwa, inconsistent_casual)
• count (int): how many times this exact (roman, devanagari) pair was independently generated across personas/batches -- a rough natural-frequency signal for how common that particular typo pattern is

Data Instance

{"roman": "namste", "devanagari": "नमस्ते", "kind": "word", "persona": "fast_thumbs", "count": 2}

Dataset Statistics

• Rows (unique roman->devanagari pairs): 858,371
• Raw variant strings generated before dedup: 893,997
• Source records (seed x persona batches): 189,941
• Kind split: {'word': 431859, 'phrase': 426512}
• Persona split: {'fast_thumbs': 430929, 'phonetic_confusion': 141389, 'inconsistent_casual': 135274, 'dropped_schwa': 150779}
• Approx. generation cost: ₹136.51 (Sarvam-30b, pay-as-you-go pricing)

Intended Use

This is augmentation training data, not a clean benchmark -- it is deliberately noisy by design so a downstream model learns robustness to sloppy typing. For clean evaluation, pair this with Aksharantar or the Dakshina dataset as a held-out test set.

How to Load

from datasets import load_dataset
ds = load_dataset("IB-Emper/hinglish-translit-sarvam")
print(ds["train"][0])

Limitations

• Synthetic data inherits any systematic biases/errors of Sarvam-30b's romanization "instincts."
• Not manually human-verified at scale; spot-checked during construction.
• Persona distribution may not be perfectly even (see stats above) if generation was interrupted and resumed.

License / Citation

Generated using the Sarvam AI API (sarvam-30b). Seed vocabulary sourced from the Leipzig Corpora Collection (CC BY). Please review Sarvam's terms of service for any commercial use of model-generated content.