To improve the translation capabilities of an OpenAI Model the fine-tuning can be used to instruct the model for specific translation needs. To achieve this, a general instruction (called "system message" ) and at least 10 examples can be used to train the selected model permanently. The used language for system-messages is English, the examples represent a translation from the configured source- to the configured target language. Additionally other language-resources like terminology and translation memories can be used to train the model. Keep in mind, that only special resources tailored for training shall be used and it makes no sense to add a general TM or terminology. For a successful training at least 10 and up to 50 examples must be added.
In the fine-tuning window the system message can directly be tested by translating a sample target sentence using the tested system message. Keep in mind this example translation will test the system message only and cannot take the 10-50 examples into account and usually the result of the training is better than what the example translation suggests.