| """ |
| LatticeSpike: Geometric Grounding and Spiking Dynamics for |
| Real-Time Hallucination Detection in LLMs. |
| |
| A reference PyTorch implementation of the architecture described in |
| `lattice_snn_hallucination.tex`. |
| |
| Modules |
| ------- |
| surrogate : Surrogate-gradient spike functions (sigmoid / arctan). |
| boxes : Box-embedding encoder, Galois concretization, geometric losses. |
| signals : Six input-channel computation + MRL projection of the KL trace. |
| rsnn : Box-volume-modulated recurrent LIF spiking network. |
| detector : End-to-end LatticeSpike module (encoding + dynamics + detection). |
| losses : Combined training objective. |
| metrics : AUPRC and lead-time evaluation. |
| llm_signals : Optional extraction of KL trace / attention / hidden states from an HF LLM. |
| """ |
|
|
| from .surrogate import SpikeFunction, sigmoid_surrogate, atan_surrogate |
| from .boxes import BoxEncoder, BoxBatch, geometric_signals, containment_loss, volume_loss, anchored_contraction_loss, box_intersection, join_loss |
| from .signals import MRLProjection, ChannelStack, assemble_channels |
| from .rsnn import VolumeModulatedRSNN |
| from .lstm_dyn import LSTMDynamics |
| from .detector import LatticeSpike, LatticeSpikeOutput |
| from .losses import ( |
| CombinedLoss, |
| LossWeights, |
| margin_ranking_anomaly_loss, |
| margin_ranking_anomaly_loss_from_sequences, |
| ) |
| from .metrics import auprc, lead_time |
|
|
| __all__ = [ |
| "SpikeFunction", "sigmoid_surrogate", "atan_surrogate", |
| "BoxEncoder", "BoxBatch", "geometric_signals", |
| "containment_loss", "volume_loss", "anchored_contraction_loss", |
| "box_intersection", "join_loss", |
| "MRLProjection", "ChannelStack", "assemble_channels", |
| "VolumeModulatedRSNN", |
| "LSTMDynamics", |
| "LatticeSpike", "LatticeSpikeOutput", |
| "CombinedLoss", "LossWeights", |
| "margin_ranking_anomaly_loss", "margin_ranking_anomaly_loss_from_sequences", |
| "auprc", "lead_time", |
| ] |
|
|
| __version__ = "0.1.0" |
|
|
