pandakingpunc/megazeka-tr-spellfix-pairs

Megazeka · Turkish Spelling Correction Pairs

44,643 (misspelled → correct) Turkish sentence pairs with a full record of every corruption that was applied, generated from the CC0 Common Voice Turkish Sentence Collector and from project-authored everyday first/second-person sentences.

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 versus chat-style spellings like gidicem? — rather than only in aggregate.

Used to train pandakingpunc/megazeka-byt5-tr-spellfix v3 (live demo: spaces/pandakingpunc/megazeka). Author: Mustafa Karatüm. Generation code: github.com/pandakingpunc/megazeka (training/prepare.py, training/everyday.py, src/megazeka/noise.py, src/megazeka/colloquial.py, src/megazeka/spelling.py).

What's new in v3

main is now v3; the previous release stays available under tag v2 (and v1).

• Four new noise operations: sert_ünsüz (hardened consonants such as gitcez/etcem, where the correct form has a softened d), ne_kısalt (napcam → ne yapacağım…), yazım_yanlışı (~35 common Turkish misspellings: yanlız, herkez, yalnış, şöför, orjinal, süpriz, makina, hastahane, yada, …) and özel_ad_kesme (Ankaraya for Ankara'ya).
• Casing bug fixed: str.swapcase turned İ into two code points and I into i; the generator now uses Turkish-aware swapcase_tr / upper_tr.
• More everyday sentences: 1,470 → 2,561 generated (2,526 used after de-duplication against Common Voice).
• Train is now 42,075 pairs (31,996 Common Voice, 8,804 everyday, 1,275 two-sentence chunks); validation was regenerated and has 1,440 pairs (640 Common Voice, 800 everyday).
• Test split frozen from v2: the 1,128 test pairs are byte-identical to v2 (SHA-256 43738c617c498e21753f71f9af928ac518c374e938742901a443e81b827e30a5), so models from v2 and v3 can be compared directly. Test targets are excluded from train and validation.
• The 81 targets of a hand-written challenge set (in the GitHub repo, evaluation/challenge.jsonl) are also excluded from train and validation. That set was written alongside the new noise rules, so it measures whether targeted failures got fixed, not general accuracy.

What was new in v2

(Kept for history.)

v1 (tag v1) had 33,280 pairs from Common Voice only. Those sentences are literary and mostly third person — gideceğim appeared 8 times in 32,000 training pairs — so a model trained on them never learned colloquial first-person forms. v2 adds:

• Everyday sentences (CC0-1.0, written for this project): template sentences from a 50-verb table and a small conjugator (future, present continuous, negative, question suffix; vowel harmony and the git/et/ye/de special cases) plus ~250 hand-written short sentences, ~1,470 in total.
• Suffix-based colloquial noise, applied to any verb rather than a fixed word list (-eceğim → -icem/-cem, -yorum → -yom, gideceğim → gitcem, burada → burda, bir şey → bişey, …). The gündelik operation count in train went from 1,001 to 8,178.
• Two-sentence chunks in train: two independent pairs of the same difficulty joined with a space, at most 170 bytes.
• Larger validation and test splits (1,128 pairs each) with an everyday subset.

The 160 Common Voice test target sentences are the same as in v1, but the noise generator changed, so most corrupted inputs differ.

Splits

Split	Pairs	Distinct targets	Common Voice	Everyday	Two-sentence
train	42,075	10,200	31,996	8,804	1,275
validation	1,440	360	640	800	—
test (frozen from v2)	1,128	282	640	488	—

Splitting happens on canonical target form (NFC + Turkish lowercase + alphanumeric key) before noise is applied, so no variant of the same sentence crosses splits. Of the 2,526 everyday sentences used, about 1/12 go to validation and the rest to train; the test split is not regenerated but carried over from v2. Seed 42.

Schema

{
  "target": "Yarın sabah ben evde kalıyorum.",
  "source": "Megazeka gündelik şablon / CC0-1.0",
  "input": "Yarın sabah benevde kalıyom.",
  "difficulty": "orta",
  "operations": [
    {"operation": "gündelik",  "before": "Yarın sabah ben evde kalıyorum.",
                               "after":  "Yarın sabah ben evde kalıyom."},
    {"operation": "boşluk_sil", "before": "Yarın sabah ben evde kalıyom.",
                               "after":  "Yarın sabah benevde kalıyom."}
  ],
  "group_id": "5864952e584f92f7"
}

• 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
• source — Common Voice / CC0-1.0, Megazeka gündelik şablon / CC0-1.0 (everyday) or Birleşik iki cümle (two-sentence chunk)
• group_id — canonical target hash; shared by the four variants of one sentence
• group_id_second — only on two-sentence chunks: the group_id of the second sentence
• operations — ordered list of applied corruptions with before/after text; on two-sentence chunks the operations of both halves are listed, each with the text of its own half

Difficulty levels are balanced (≈10,500 pairs each in train). 25% of pairs are clean, so a model trained on this data learns to leave correct text alone. For easy and medium everyday pairs the first operation is always gündelik. Rare multi-operation corruptions can cancel each other out — when measuring overcorrection, count every example where input == target (10,569 in train, 282 in test) rather than trusting the temiz label.

Noise operations

Counts over the training split (v3), including two-sentence chunks (a pair may apply several):

Operation	Count	What it does
türkçe_karakter	15,480	Diacritic loss: ı→i, ş→s, ğ→g, ç→c, ö→o, ü→u
gündelik	8,853	Colloquial spellings (gidicem, yapcaz, bilmiyom, burda, bişey, …)
küçük_harf	8,084	Capitalization errors (Turkish-aware as of v3)
harf_sil	4,688	Dropped letter
boşluk_ekle	4,567	Split word
boşluk_sil	4,545	Merged words
klavye	4,524	Turkish Q-keyboard neighbour key
noktalama_ekle	4,517	Spurious punctuation
harf_tekrar	4,483	Repeated letter
fazla_boşluk	4,482	Extra whitespace
noktalama_sil	4,352	Missing punctuation
harf_yer_değiştir	3,757	Transposed adjacent letters
yazım_yanlışı	991	New in v3. Common misspellings (yanlız, herkez, yalnış, şöför, makina, …)
özel_ad_kesme	729	New in v3. Dropped apostrophe before a suffix on a proper noun (Ankaraya)
ne_kısalt	478	New in v3. ne yap- contractions (napcam, napıyosun)
sert_ünsüz	449	New in v3. Hardened consonant (gitcez, etcem)
soru_eki	389	Question suffix wrongly attached
de_da	219	de/da clitic merged with the word
ki	156	ki clitic merged with the word
kesme	114	Missing apostrophe before a suffix

Source and filtering

• Common Voice Turkish Sentence Collector 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.
• Everyday sentences, generated by training/everyday.py in the GitHub repository and released here under CC0-1.0 (2,561 generated in v3, 2,526 used).

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.
• The everyday sentences come from 50 verbs and a limited set of complements; they do not represent real chat language in full, and the test everyday subset is from the same template family as train.
• Common Voice sentences may themselves contain overlooked mistakes; passing the filter is not evidence of grammatical correctness.
• Sentence-length only (plus two-sentence chunks). No long documents, lists, code, or mixed-language text.
• Split separation is on canonical sentence form. Semantically similar sentences — including template sentences that share a verb or complement — may still be split across sets, and overlap with any given model's pretraining corpus is unknown.

Versions

Version	Revision	Pairs (train / val / test)	Sources
v3	main	42,075 / 1,440 / 1,128 (test frozen from v2)	Common Voice + more everyday + two-sentence chunks, 4 new noise operations
v2	tag v2	38,057 / 1,128 / 1,128	Common Voice + everyday + two-sentence chunks
v1	tag v1	32,000 / 640 / 640	Common Voice only

The default is v3. Load a specific version with load_dataset("pandakingpunc/megazeka-tr-spellfix-pairs", revision="v2") (or "v1").

Licensing

Common Voice source sentences are CC0-1.0 (Common Voice licensing); the everyday sentences were written for this project and are also released under CC0-1.0, as are all derived pairs. No Common Voice program code (MPL-2.0) is included here. The generation code is MIT, in the GitHub repository.