BaseAlgorithmConfig

Fields common to every training algorithm.

Concrete algorithms subclass this and add their own knobs (SDFT: topk; RL: num_samples / verifier_name / …). extra="forbid" so a YAML typo fails loudly. A subclass may re-declare a field to change its default (e.g. RL drops learning_rate to 1e-5).

Attributes

attributemodel_config

= ConfigDict(extra='forbid')

attributelearning_ratefloat

= 0.0001

attributenum_epochsint

= 1

attributebatch_sizeint

= Field(default=4, gt=0)

attributemax_stepsint | None

= None

Hard step cap. When set, wins over num_epochs * steps_per_epoch.

attributelora_rankint

= 8

attributerenderer_namestr | None

= None

attributeenable_thinkingbool | None

= None

attributesave_everyint

= 0

If 0, computed from save_at_fractions (GCD-of-marks heuristic).

attributesave_at_fractionslist[float]

= Field(default_factory=(lambda: [1.0]))

attributecallbackslist[CallbackSpec]

= Field(default_factory=list)

attributeadam_beta1float

= 0.9

attributeadam_beta2float

= 0.95

attributeadam_epsfloat

= 1e-08

attributewandb_projectstr | None

= None

attributewandb_namestr | None

= None