API Reference#

Model API#

api.ICEICLModel

In-Context Learning model with ICE (Identification–Conditioning–Equalization) backbone.

Factory & Registry#

model.make_model

Construct a model by arm name.
iceicl.model.ARM_REGISTRY: dict[str, type[torch.nn.Module]]#

Registry mapping arm name strings to model classes. Twelve entries, one per ablation arm:

Name

Class

Params

baseline

BaselineNanoICL

27.6 M

cross_attn

CrossAttnICL

27.6 M

ssm_cross

HybridSSMICL

12.6 M

ssm_cross_fb

HybridSSMICL_Feedback

12.6 M

fb_pseudo_pilot

HybridSSMICL_FB_PseudoPilot

12.6 M

fb_incremental

HybridSSMICL_FB_Incremental

12.9 M

fb_confidence

HybridSSMICL_FB_Confidence

12.6 M

fb_turbo

HybridSSMICL_FB_Turbo

12.6 M

fb_innovation

HybridSSMICL_FB_Innovation

12.9 M

fb_iter2

HybridSSMICL_FB_Iter2

12.9 M

direct_tau

HybridSSMICL_DirectTau

12.9 M

fb_anti_conf

HybridSSMICL_FB_AntiConf

12.9 M

Use make_model() for convenient construction by name.

Utilities#

utils.calibration_curve

Compute reliability-diagram data.

utils.expected_calibration_error

Expected calibration error (ECE).