| --- |
| license: mit |
| language: |
| - tr |
| size_categories: |
| - 10K<n<100K |
| --- |
| |
| # Turkish Pseudo-Reasoning Dataset |
|
|
| This project creates synthetic reasoning traces in Turkish for simple, fully |
| verifiable tasks. Its goal is to provide a pseudo-reasoning dataset with 100% |
| label accuracy, suitable for establishing a controlled reasoning baseline |
| before experimenting with noisier, model-generated, or open-ended traces. |
|
|
| ## Topics |
|
|
| The current dataset covers: |
|
|
| - Addition |
| - Subtraction |
| - Multiplication, including signed numbers |
| - Turkish letter counting |
| - Unit conversion across volume, length, and mass units |
|
|
| ## Methodology |
|
|
| Each example is generated from deterministic task logic and rendered through |
| varied Turkish question, reasoning, and answer templates. The traces explain |
| intermediate operations such as carrying, borrowing, sign handling, character |
| enumeration, and conversion equations. Turkish decimal notation, grammatical |
| suffixes, unit abbreviations, zero values, and task-specific edge cases are |
| handled explicitly. |
|
|
| Answers are calculated by code rather than produced by a language model. The |
| generation script enforces unique prompts per task type, checks expected row |
| counts, writes the dataset to Parquet, reads it back, and runs consistency |
| checks on randomly sampled records. Conversion records additionally use |
| deterministic equations and bounded decimal formatting. |
|
|
| ## Intended Use |
|
|
| The dataset is intended for supervised fine-tuning, evaluation, ablation |
| studies, and baseline experiments involving Turkish reasoning traces. |
