pandakingpunc/megazeka-tr-spellfix-pairs

Megazeka · Turkish Spelling Correction Pairs

33,280 (misspelled → correct) Turkish sentence pairs with a full record of every corruption that was applied, generated from the CC0 Common Voice Turkish Sentence Collector.

The distinguishing feature is the operations field: each pair carries the exact sequence of noise transformations that produced it, with before/after text at each step. That makes it possible to evaluate a corrector per error type — how well does it handle diacritic loss versus merged words? — rather than only in aggregate.

Used to train pandakingpunc/megazeka-byt5-tr-spellfix. Generation code: github.com/pandakingpunc/megazeka (training/prepare.py).

Splits

Split	Pairs	Distinct targets
train	32,000	8,000
validation	640	160
test	640	160

Splitting happens on canonical target form (NFC + Turkish lowercase + alphanumeric key) before noise is applied, so no variant of the same sentence crosses splits. Seed 42.

Schema

{
  "target": "Bir sürgü çekildi, kadın kapıyı açtı.",
  "input": "Bir surgu ckeildi kadin kapiyi acti",
  "difficulty": "zor",
  "group_id": "a1b2c3d4e5f60718",
  "source": "Common Voice / CC0-1.0",
  "operations": [
    {"operation": "türkçe_karakter", "before": "Bir sürgü çekildi, kadın kapıyı açtı.",
                                      "after": "Bir surgu cekildi, kadin kapıyi acti."},
    {"operation": "noktalama_sil",   "before": "Bir surgu cekildi, kadin kapiyi acti.",
                                      "after": "Bir surgu cekildi kadin kapiyi acti"},
    {"operation": "harf_yer_değiştir","before": "Bir surgu cekildi kadin kapiyi acti",
                                      "after": "Bir surgu ckeildi kadin kapiyi acti"}
  ]
}

• target — correct sentence
• input — corrupted sentence, identical to target when difficulty is temiz
• difficulty — temiz (clean) / kolay (easy) / orta (medium) / zor (hard), targeting 0 / 1 / 2 / 4 successfully applied operations
• group_id — canonical target hash; shared by the four variants of one sentence
• operations — ordered list of applied corruptions with before/after text

Difficulty levels are evenly balanced: 8,000 pairs each in train. 25% of pairs are clean, so a model trained on this data learns to leave correct text alone. Note that rare multi-operation corruptions can cancel each other out — when measuring overcorrection, count every example where input == target rather than trusting the temiz label.

Noise operations

Counts over the training split (a pair may apply several):

Operation	Count	What it does
türkçe_karakter	13,292	Diacritic loss: ı→i, ş→s, ğ→g, ç→c, ö→o, ü→u
küçük_harf	7,086	Capitalization errors
boşluk_sil	3,927	Merged words
harf_tekrar	3,897	Repeated letter
boşluk_ekle	3,894	Split word
fazla_boşluk	3,870	Extra whitespace
klavye	3,865	Turkish Q-keyboard neighbour key
harf_sil	3,841	Dropped letter
noktalama_ekle	3,833	Spurious punctuation
noktalama_sil	3,563	Missing punctuation
harf_yer_değiştir	3,177	Transposed adjacent letters
gündelik	1,001	Informal spellings (gidicem, yapcaz, …)
soru_eki	279	Question suffix wrongly attached
de_da	239	de/da clitic merged with the word
ki	120	ki clitic merged with the word
kesme	116	Missing apostrophe before a suffix

Source and filtering

Single source: the Common Voice Turkish Sentence Collector file at pinned commit b12cde653757295103c5948c1cd601b166e1b905 (1,557,129 bytes, 47,781 lines, SHA-256 28e2d095…4058ece). No audio or dev corpus was downloaded.

Sentences were kept if they are 3–24 words, 18–160 UTF-8 bytes, start with a capital letter and end with punctuation, then deduplicated on a canonical key. Per-file SHA-256 checksums for the generated splits are in metadata.json.

Limitations

• Synthetic. These are programmatically corrupted sentences, not errors real people made. A model that scores well here may do noticeably worse on genuine user text.
• Operation frequencies reflect the generator's weights, not measured Turkish error frequencies.
• Common Voice sentences may themselves contain overlooked mistakes; passing the filter is not evidence of grammatical correctness.
• Sentence-length only. No long documents, lists, code, or mixed-language text.
• Split separation is on canonical sentence form. Semantically similar sentences may still be split across sets, and overlap with any given model's pretraining corpus is unknown.

Licensing

Source sentences are CC0-1.0 (Common Voice licensing), and the derived pairs are released under CC0-1.0 as well. No Common Voice program code (MPL-2.0) is included here. The generation code is MIT, in the GitHub repository.