Fine-tuning
İnce ayarlama
D2
Fine-tuning is further training that adapts a pre-trained model for a specific task or domain using task-specific data.
Review status: 2026-11-26
Technical explanation
It follows pre-training and can specialize a broadly trained model for a defined application task, often through supervised learning.
Conceptual boundaries
Fine-tuning further trains a pre-trained model; it is not training from scratch, changing a prompt, or retrieving information at inference time.
Provider-neutral example
A team can further train a pre-trained model on reviewed support categories to specialize it for routing requests in one department.
Limitations
Fine-tuning can affect safety and security controls, so its effects should be evaluated in the intended system context.
Related concepts
Atomic claims and evidence
1.1NIST AI 100-2e2025, Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations
- Source
- NIST AI 100-2e2025, Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations
- Source role
- Authoritative source
- Exact locator
- Appendix A p. 108: fine-tuning
- Supported claim
- Fine-tuning adapts a pre-trained model to specific tasks or a particular domain through further training on task-specific data.
- Last verification
- Review due
- Scope limitation
- The cited definition says this is often supervised learning, not that every fine-tuning process is supervised.
2.1NIST AI 600-1, Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile
- Source
- NIST AI 600-1, Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile
- Source role
- Authoritative source
- Exact locator
- Measure p. 30, MS-2.7-008: verify fine-tuning does not compromise controls
- Supported claim
- NIST recommends verifying that fine-tuning does not compromise safety and security controls.
- Last verification
- Review due
- Scope limitation
- This is risk-management guidance and not evidence that every fine-tuned model compromises controls.
3.1NIST AI 100-2e2025, Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations
- Source
- NIST AI 100-2e2025, Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations
- Source role
- Authoritative source
- Exact locator
- Section 3.1.1, p. 37, Figure 4: LLM enterprise adoption pipeline
- Supported claim
- NIST’s LLM adoption pipeline distinguishes fine-tuning from prompt engineering and from application integration such as RAG.
- Last verification
- Review due
- Scope limitation
- This distinguishes workflow stages and does not establish a particular outcome from any one stage.