FFT
FFT (Full Fine-Tuning) updates all model weights during training. This provides the deepest level of customization but requires more compute resources and longer training time.
When to use FFT
Large domain shifts where the model needs fundamental knowledge changes
You have a large, high-quality dataset (small datasets risk catastrophic forgetting)
You want no adapter overhead at inference time
Maximum customization depth is required
Key differences from LoRA
No
lora_rank,lora_alpha, ormerge_weightsparametersAdds
max_response_lengthparameter for controlling generated output lengthHigher compute cost and longer training time
Larger output artifacts (full model weights)
Not compatible with continuous customization
For the full parameter reference including FFT-specific hyperparameters, see each technique page: SFT, DPO, RFT.
Dataset compatibility
FFT uses the same dataset formats as LoRA for each customization technique. You do not need a separate dataset schema when you switch from LoRA to FFT.
SFT FFT uses the same dataset formats described in Dataset format.
DPO FFT uses the same dataset formats described in Dataset format.
RLVR FFT uses the same dataset format described in Dataset format.
RLAIF FFT uses the same dataset format described in Dataset format.
The main difference between LoRA and FFT is how weights are updated and which hyperparameters are available, not how you structure your training data.