| --- |
| license: cc-by-4.0 |
| task_categories: |
| - translation |
| language: |
| - hi |
| tags: |
| - transliteration |
| - romanization |
| - hindi |
| - hinglish |
| - input-method |
| size_categories: |
| - 100K<n<1M |
| --- |
| |
| # Hinglish -> Hindi Transliteration Dataset (sloppy roman input) |
|
|
| Parallel data for training a **romanized-Hindi -> Devanagari** transliteration |
| model that tolerates the messy, inconsistent way people actually type Hindi on |
| phones. Built to power a smart Hindi input method (IME), analogous to Chinese |
| smart-pinyin engines. |
|
|
| Each row is one `(roman, devanagari)` pair: `roman` is a plausible sloppy |
| human romanization, `devanagari` is the correct standard Hindi. |
|
|
| ## Columns |
| - `roman` — model input, sloppy romanized Hindi |
| - `devanagari` — model target, correct Devanagari |
| - `count` — occurrences of this exact pair across generation (common typo |
| patterns have higher counts; usable as a sampling weight) |
| - `mode` — provenance stream (see below) |
| - `persona` / `register` — typing-style and code-switch metadata (nullable) |
|
|
| ## Provenance (`mode`) |
| - **chatroman** — realistic WhatsApp-style conversations generated in |
| Devanagari, then romanized via a numbered ID-echo protocol so the |
| Devanagari is never model-invented. |
| - **sentence** — real sentences from the Leipzig `hin_news_2011_1M` corpus |
| (CC BY), romanized via the same ID-echo protocol. Ground-truth Hindi. |
| - **surgical** — a hand-curated priority vocabulary (loanwords, confusable |
| near-homophones, high-frequency function words, kinship/number/date terms) |
| with many romanization variants each, to anchor the hardest short-word cases. |
| - **scenario** — legacy single-pass conversational data (small remainder). |
|
|
| Romanizations were produced by Sarvam's Indic LLM across 10 typing personas |
| (fast-thumbs, dropped-schwa, gen-z abbreviator, boomer-formal, etc.) and 3 |
| registers (pure Hindi, light/heavy code-switch). Correct Hindi ground truth |
| comes only from the Leipzig corpus and the curated list; the LLM authored |
| casual conversational Hindi for the chatroman stream, validated for script and |
| structure. About 13% of rows retain in-line English words on the Devanagari |
| side (real code-switch, kept intentionally). |
|
|
| ## Splits |
| `train` / `validation` (validation is a small random holdout for loss tracking; |
| it is NOT a clean transliteration benchmark — use Dakshina or Aksharantar for |
| held-out evaluation). |
|
|
| ## Cleaning |
| Line-level validation during generation: script/charset checks, digit and |
| garbage rejection, ID-echo word-count parity (anti-invention), and per-mode |
| quarantine with reason codes. Exact-duplicate pairs collapsed with a `count`. |
|
|
| ## Intended use |
| Training small char-level seq2seq transliteration models for Hindi IMEs. |
| Not a general MT dataset. |
|
|
