Keigo_0
It achieves the following results on the evaluation set:
Model description
The model is designed for sentence-level rewriting, specifically transforming rude, informal, aggressive, or
slang-heavy Turkish sentences into polite, formal, and respectful expressions, while preserving the original meaning.
Keigo_0 focuses on tone transformation, not content moderation or censorship.
The goal is to provide a culturally appropriate, professional rephrasing aligned with formal Turkish usage.
Intended uses & limitations
Intended uses
-Rewriting rude, aggressive, or slang Turkish sentences into:
Formal - Polite - Professional tone
-Improving user-facing messages in:
Applications - Customer support systems - Feedback forms
Limitations
-The model is not a chatbot and should not be used for open-ended conversation.
-It is optimized for short, single-sentence inputs; long paragraphs may produce degraded results.
-Output quality depends on input clarity; ambiguous or incomplete inputs may lead to generic responses.
Training and evaluation data
The model was fine-tuned on a custom-curated Turkish dataset consisting of paired sentence examples.
Each training sample contains:
Input: An informal, rude, slang-heavy, or emotionally aggressive Turkish sentence
Target: A rewritten version of the same sentence in a formal, polite, and professional tone, while preserving the original meaning
Data characteristics
Language: Turkish
Format: Input–target sentence pairs
Style focus:
Tone softening
Formalization
Politeness transformation
Sentence length: Short, single-sentence inputs
Domain:
Everyday user complaints
Informal speech
Emotional or frustrated expressions
Data sourcing
The dataset was manually created and curated for this project.
No personally identifiable information (PII) was intentionally included.
-Evaluation-
Evaluation was performed using:
A held-out validation split from the same dataset
Cross-entropy loss as the primary metric
The evaluation focuses on:
Meaning preservation
Tone transformation quality
Fluency and grammatical correctness
Training procedure
Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 0.0004
- train_batch_size: 4
- eval_batch_size: 8
- seed: 42
- gradient_accumulation_steps: 2
- total_train_batch_size: 8
- lr_scheduler_type: linear
- lr_scheduler_warmup_steps: 100
- num_epochs: 40
Training results
training loss=0.064679
eval loss: 0.209111
Framework versions
- Transformers 4.57.3
- Pytorch 2.9.1+cu128
- Datasets 4.4.2
- Tokenizers 0.22.1
