LocalSFTConfig

Attributes

attributemodel_config

= ConfigDict(extra='forbid')

attributelearning_ratefloat

= 0.0002

attributenum_epochsint

= 3

attributemax_stepsint | None

= None

If set, overrides num_epochs - training stops after this many steps.

attributeper_device_train_batch_sizeint

= 1

attributegradient_accumulation_stepsint

= 16

attributewarmup_stepsint

= 20

attributemax_seq_lenint

= 512

attributelogging_stepsint

= 10

attributesave_stepsint

= 100

attributesave_total_limitint

= 5

attributebf16bool

= False

Enable only on CUDA GPUs that support bfloat16 (Ampere+). Crashes on CPU/MPS.

attributefp16bool

= False

Enable only on CUDA GPUs. Crashes on CPU/MPS.

attributeseedint

= 42

attributelora_rankint

= 16

attributelora_alphaint

= 32

attributelora_dropoutfloat

= 0.05

attributelora_target_moduleslist[str]

= Field(default_factory=(lambda: ['q_proj', 'k_proj', 'v_proj', 'o_proj', 'gate_proj', 'up_proj', 'down_proj']))