| --- |
| language: |
| - en |
| - tr |
| task_categories: |
| - translation |
| pretty_name: Tokenized English-Turkish Translation Corpus |
| configs: |
| - config_name: default |
| default: true |
| data_files: |
| - split: train |
| path: "data/train-*.parquet" |
| --- |
| |
| # Tokenized English-Turkish Translation Corpus |
|
|
| Format-augmented and tokenized subset of `Ba2han/finetranslations-TR_filtered`, built with |
| `Ba2han/outputs`. |
|
|
| ## Token amounts |
|
|
| | Metric | Amount | |
| |---|---:| |
| | Target tokens | 500,000,000 | |
| | Actual tokens | 500,000,186 | |
| | Full-example overshoot | 186 | |
| | Output rows | 343,740 | |
| | Average tokens/row | 1,454.59 | |
| | Minimum tokens/row | 441 | |
| | Maximum tokens/row | 5,472 | |
| | Source rows scanned | 204,362 | |
| | Source rows passing filters | 42,976 | |
| | Parquet shards | 4 | |
| | Build time | 0.10 hours | |
|
|
| The script stops after the first complete example that reaches the target, so |
| no example is cut merely to make the total exactly 500,000,000 tokens. |
|
|
| ## Filters |
|
|
| ```python |
| total_token_count > 800 |
| og_quality_score > 0.05 |
| ``` |
|
|
| `translated_text` is formatted as English and `og_full_text` as Turkish, as |
| specified for this build. |
|
|
| ## Output columns |
|
|
| - `input_ids`: token IDs from `Ba2han/outputs` |
| - `num_tokens`: length of `input_ids` |
| - `augmentation_id`: formatting variant from 0 through 7 |
|
|
| Tokenization uses `add_special_tokens=True`, with no padding and no truncation. |
| An attention mask is not stored because every saved sequence is unpadded. |
|
|
| ```json |
| { |
| "bos_token": "<bos>", |
| "bos_token_id": 2, |
| "eos_token": "<eos>", |
| "eos_token_id": 3, |
| "pad_token": "<pad>", |
| "pad_token_id": 0 |
| } |
| ``` |
|
|
| ## Tokens by augmentation |
|
|
| | ID | Template | Rows | Tokens | Average | Min | Max | |
| |---:|---|---:|---:|---:|---:|---:| |
| | 0 | `English: ...\nTurkish: ...` | 42,965 | 62,451,999 | 1,453.56 | 442 | 5,469 | |
| | 1 | `Türkçe: ...\n\nEnglish: ...` | 42,973 | 62,460,327 | 1,453.48 | 442 | 5,469 | |
| | 2 | `English:\n...\n\nTürkçe:\n...` | 42,966 | 62,419,132 | 1,452.76 | 441 | 5,468 | |
| | 3 | `Turkish:\n...\n\nEnglish:\n...` | 42,966 | 62,454,483 | 1,453.58 | 442 | 5,469 | |
| | 4 | `[English]\n...\n\n[Turkish]\n...` | 42,966 | 62,556,486 | 1,455.95 | 444 | 5,471 | |
| | 5 | `[Türkçe]\n...\n\n[English]\n...` | 42,969 | 62,528,589 | 1,455.20 | 443 | 5,470 | |
| | 6 | `English text:\n...\n\nTurkish text:\n...` | 42,970 | 62,573,215 | 1,456.21 | 445 | 5,472 | |
| | 7 | `Türkçe metin:\n...\n\nİngilizce metin:\n...` | 42,965 | 62,555,955 | 1,455.97 | 444 | 5,471 | |
|
|
| ## Tokens by shard |
|
|
| | Shard | File | Rows | Tokens | Compressed size | |
| |---:|---|---:|---:|---:| |
| | 0 | `train-00000-of-00004.parquet` | 100,000 | 145,603,263 | 61.85 MiB | |
| | 1 | `train-00001-of-00004.parquet` | 100,000 | 145,557,942 | 61.85 MiB | |
| | 2 | `train-00002-of-00004.parquet` | 100,000 | 144,621,306 | 61.32 MiB | |
| | 3 | `train-00003-of-00004.parquet` | 43,740 | 64,217,675 | 27.32 MiB | |
|
|
| ## Load |
|
|
| ```python |
| from datasets import load_dataset |
| |
| ds = load_dataset("Ba2han/tokenized_translation", split="train") |
| ``` |
|
|
