DUSUNEN-Nokta-68M-v1

DUSUNEN Nokta 68M v1

One compact Turkish model for punctuation restoration and true-casing. Nokta predicts both decisions in a single pass through 28 joint token labels, so deployment does not require a punctuation model followed by a second casing model.

Held-out results

The split is deterministic and leakage-resistant: all rows sharing a source URL remain in the same 98/1/1 train/validation/test bucket. The test bucket was not used for gradient updates or checkpoint selection.

Punctuation token accuracy is 96.57% and casing token accuracy is 97.00%. Machine-readable results and class counts are included in this repository. dataset-audit.json records file checksums, label distributions and zero source overlap across splits.

Try it

The repository includes inference.py, which handles Turkish dotted/dotless I, subword-to-word alignment, punctuation marks and long-text chunking:

python inference.py   "bugün istanbulda hava güzel mi yarın yağmur bekleniyor"   --model GoktugD/DUSUNEN-Nokta-68M-v1

The checkpoint also loads directly with the standard Transformers API:

from transformers import AutoModelForTokenClassification, AutoTokenizer

model_id = "GoktugD/DUSUNEN-Nokta-68M-v1"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForTokenClassification.from_pretrained(model_id)

Validated ONNX

onnx/model.onnx is a lossless FP32 CPU export (270.8 MB). Across 64 held-out examples, its token prediction agreement with PyTorch was 100.000%. The recorded single-item PyTorch/ONNX median-latency ratio was 0.65×; this release makes no ONNX speedup claim. See onnx-validation.json for raw latency and numerical-parity values. A separately tokenized variable-length input also achieved 100.000% prediction agreement.

Training and data

• Base model: dbmdz/distilbert-base-turkish-cased (MIT), 68.1M parameters
• Corpus: Cosmos Turkish Corpus v1.0 (CC BY 4.0)
• Examples: 146,936 train / 1,519 validation / 1,545 test
• Input length: at most 128 subword tokens
• Optimisation: AdamW, linear schedule, BF16, seed 42
• Hardware: NVIDIA GeForce RTX 3090 24 GB
• Training time: 725.7 seconds
• Exact programs: prepare_nokta_dataset.py, train_nokta.py, export_nokta_onnx.py

Intended use and limitations

Nokta is intended for cleaning ASR output, subtitle drafts, search transcripts, chat text and Turkish document preprocessing. It was trained on web text rather than spontaneous speech transcripts, so domain punctuation can differ. Proper names, stylised brands, mixed-language passages, URLs and deliberately unusual casing remain difficult. The MIXED label restores acronym stems before an apostrophe but cannot reconstruct every arbitrary character-level casing pattern. Human review is required for legal, medical or other high-stakes documents.

Integrity

• model.safetensors SHA-256: ea24882907321293322bd2262e9684d16a2c45ff7f2a6bb89c579b8270150d5b
• onnx/model.onnx SHA-256: 0444bd6786277999d4ce6612693d4e41ec56fffabd654441a04d1c562b2dc798

Developed and released by Göktuğ Düşünen.