Synthetic data
Sentetik veri
D1 · Foundations, models and data
Synthetic data is artificially generated data used in place of or alongside directly collected observations.
Review status: 2026-12-05
Technical explanation
In language-model training, generated instructions and responses can be filtered and used as training examples.
Conceptual boundaries
Generation describes origin; it does not by itself establish data quality or suitability for a task.
Provider-neutral example
Illustrative: a team generates draft support questions, checks them and keeps suitable ones for a training dataset.
Limitations
Repeated training on generated outputs can degrade coverage of the original data distribution in studied settings.
Related concepts
Atomic claims and evidence
1.1Self-Instruct: Aligning Language Models with Self-Generated Instructions
- Source
- Self-Instruct: Aligning Language Models with Self-Generated Instructions
- Source role
- Authoritative source
- Exact locator
- Abstract: generated instructions, inputs and outputs; filtering before fine-tuning
- Supported claim
- Self-Instruct generates instructions, inputs and outputs, then filters invalid or similar samples before fine-tuning.
- Last verification
- Review due
- Scope limitation
- This is one model-generated training-data pipeline, not all synthetic data.
2.1The Curse of Recursion: Training on Generated Data Makes Models Forget
- Source
- The Curse of Recursion: Training on Generated Data Makes Models Forget
- Source role
- Authoritative source
- Exact locator
- Abstract: model collapse in training with model-generated content
- Supported claim
- Shumailov and colleagues report model collapse when models train on generated content under their studied conditions.
- Last verification
- Review due
- Scope limitation
- This does not imply that every use or mixture of synthetic data causes collapse.