LanguageModelTrainableScope
Enums · AiDotNet.Enums
Which parts of a multimodal language model a training step may update.
For Beginners
Large language models are expensive to retrain and easy to damage with a small dataset. Many recipes keep most of the model fixed and train only the pieces that connect it to the new task. This setting chooses how much of the model learns.
Fields
| Field | Summary |
|---|---|
LanguageModelTrainableScope All |
Every parameter of the model is trainable. |
LanguageModelTrainableScope None |
The whole model is frozen and used only to produce features. |
LanguageModelTrainableScope WordEmbeddingsAndHead |
Only the word (token) embeddings and the language-model head are trainable; the vision encoder, projector, transformer layers and positional embeddings are frozen. |
int value__ |