---
language:
  - tr
tags:
  - turkish
  - corpus
  - sentencepiece
  - unigram
  - tokenized
---

# my_turkish_corpus_1

Cleaned and tokenized Turkish text corpus for language model training.

## Dataset Summary

| Property | Value |
|---|---|
| Language | Turkish |
| Token count | ~369M |
| Vocab size | 8192 (SentencePiece Unigram) |
| UNK rate | 0% |
| Chars/token | 3.62 |
| Token dtype | int32 |

## Cleaning

- Fixed 38K+ broken Latin-1→UTF8 encoding errors
- Removed Myanmar/Burmese contamination (58 lines)
- Removed HTML tags, junk lines, empty lines, adjacent duplicates
- Normalized whitespace, removed soft hyphens & non-breaking spaces

## Files

| File | Description | Size |
|---|---|---|
| `tr_unigram_8192_corpus_ids.npy` | Tokenized corpus (flat int32) | ~1.48 GB |
| `tr_unigram_8192.model` | SentencePiece Unigram model | ~374 KB |
| `tr_unigram_8192.vocab` | Vocabulary text file | ~144 KB |
| `spm_vocab.json` | Vocabulary JSON (piece→id) | ~155 KB |
| `spm_meta.json` | Training metadata | ~586 B |

## Usage

```python
import numpy as np
from huggingface_hub import hf_hub_download
import sentencepiece as spm

npy = hf_hub_download("dcx514ai/my_turkish_corpus_1", "tr_unigram_8192_corpus_ids.npy")
mdl = hf_hub_download("dcx514ai/my_turkish_corpus_1", "tr_unigram_8192.model")

ids = np.load(npy)  # ~369M tokens
sp = spm.SentencePieceProcessor(model_file=mdl)
text = sp.decode(ids[:100].tolist())
print(text)
```

## Stats

- `tr_corpus_clean.txt`: 2,029,958 lines / 696 MB
- `tr_corpus2_clean.txt`: 419,629 lines / 640 MB
- Total: 2,449,587 clean lines / ~1.34 GB raw text → 369M tokens