transformer_lens.tools package¶
Subpackages¶
- transformer_lens.tools.analysis package
- Submodules
- transformer_lens.tools.analysis.attribution_patching module
- transformer_lens.tools.analysis.backward_lens module
- transformer_lens.tools.analysis.direct_logit_attribution module
- transformer_lens.tools.analysis.direct_path_patching module
- transformer_lens.tools.analysis.jacobian_lens module
- transformer_lens.tools.analysis.jacobian_lens_causal_swap_benchmark module
AnswerMetricsBaselineRecordBenchmarkCorpusControlDrawFunctionSpecPromptTrialSpecSuccessRateIntervalTrialResultbuild_protocol_manifest()compute_answer_metrics()corpus_definition()filter_baseline_capable()fingerprint_manifest()iter_prompt_trials()load_artifact()results_fingerprint()run_causal_swap_benchmark()run_causal_swap_trial()select_displacement_matched_control_token()serialize_artifact()success_rate_ci()
- transformer_lens.tools.analysis.jacobian_lens_causal_swap_benchmark_cli module
- transformer_lens.tools.analysis.jacobian_lens_coordinate_patch module
- transformer_lens.tools.analysis.jacobian_lens_decomposition module
- transformer_lens.tools.analysis.projection_kernel module
- transformer_lens.tools.analysis.sparse_probing module
- transformer_lens.tools.analysis.svd_circuits module
- Module contents
ActivationProjectionAttentionHeadRefAttributionResultBackwardLensBackwardLensLayerResultBackwardLensMatrixResultBackwardLensMatrixResult.bottom()BackwardLensMatrixResult.bottom_rankingBackwardLensMatrixResult.bottom_tokens()BackwardLensMatrixResult.factor_normsBackwardLensMatrixResult.factorsBackwardLensMatrixResult.gradient_descent_target_ranks()BackwardLensMatrixResult.logits()BackwardLensMatrixResult.normalized_bottom_rankingBackwardLensMatrixResult.normalized_target_largest_ranksBackwardLensMatrixResult.normalized_target_smallest_ranksBackwardLensMatrixResult.normalized_top_rankingBackwardLensMatrixResult.normalized_vocabulary_logitsBackwardLensMatrixResult.projected_factorBackwardLensMatrixResult.target_largest_ranksBackwardLensMatrixResult.target_ranks()BackwardLensMatrixResult.target_smallest_ranksBackwardLensMatrixResult.target_token_idBackwardLensMatrixResult.top()BackwardLensMatrixResult.top_rankingBackwardLensMatrixResult.top_tokens()BackwardLensMatrixResult.vocabulary_logitsBackwardLensMatrixResult.vocabulary_sizeBackwardLensMatrixResult.zero_norm_mask
BackwardLensResultBackwardLensResult.includes_full_logitsBackwardLensResult.includes_normalized_logitsBackwardLensResult.layer()BackwardLensResult.layersBackwardLensResult.lossBackwardLensResult.max_absolute_reconstruction_errorBackwardLensResult.max_relative_reconstruction_errorBackwardLensResult.promptBackwardLensResult.prompt_token_idsBackwardLensResult.target_tokenBackwardLensResult.target_token_id
BenchmarkCorpusCoordinatePatchCoordinatePatch.support_beforeCoordinatePatch.support_afterCoordinatePatch.coordinates_beforeCoordinatePatch.coordinates_afterCoordinatePatch.source_slotCoordinatePatch.target_slotCoordinatePatch.target_was_appendedCoordinatePatch.target_was_selectedCoordinatePatch.overwritten_target_coordinateCoordinatePatch.reconstruction_beforeCoordinatePatch.reconstruction_afterCoordinatePatch.residualCoordinatePatch.deltaCoordinatePatch.patchedCoordinatePatch.nonedited_coordinate_max_deltaCoordinatePatch.residual_max_deltaCoordinatePatch.source_target_cosineCoordinatePatch.basis_rankCoordinatePatch.basis_condition_numberCoordinatePatch.coordinates_after_nonnegativeCoordinatePatch.basis_condition_numberCoordinatePatch.basis_rankCoordinatePatch.coordinates_afterCoordinatePatch.coordinates_after_nonnegativeCoordinatePatch.coordinates_beforeCoordinatePatch.deltaCoordinatePatch.nonedited_coordinate_max_deltaCoordinatePatch.overwritten_target_coordinateCoordinatePatch.patchedCoordinatePatch.reconstruction_afterCoordinatePatch.reconstruction_beforeCoordinatePatch.residualCoordinatePatch.residual_max_deltaCoordinatePatch.source_slotCoordinatePatch.source_target_cosineCoordinatePatch.support_afterCoordinatePatch.support_beforeCoordinatePatch.target_slotCoordinatePatch.target_was_appendedCoordinatePatch.target_was_selected
DegenerateDirectionErrorDirectLogitAttributionEdgeAttributionConfigFunctionSpecHeadAffinityPairHeadAffinityResultHeadAffinityResult.normalizedHeadAffinityResult.rankHeadAffinityResult.rtolHeadAffinityResult.scoresHeadAffinityResult.source_head_kindHeadAffinityResult.source_layer_indicesHeadAffinityResult.source_rankHeadAffinityResult.source_ranksHeadAffinityResult.source_roleHeadAffinityResult.target_head_kindHeadAffinityResult.target_layer_indicesHeadAffinityResult.target_rankHeadAffinityResult.target_ranksHeadAffinityResult.target_roleHeadAffinityResult.top_pairs()HeadAffinityResult.valid_mask
HeadDecompositionHeadSVDHeadSVD.whichHeadSVD.layerHeadSVD.headHeadSVD.UHeadSVD.SHeadSVD.VHeadSVD.rank_reportHeadSVD.epsHeadSVD.null_rtolHeadSVD.folded_lnHeadSVD.SHeadSVD.UHeadSVD.VHeadSVD.block_of()HeadSVD.degenerate_blocks()HeadSVD.epsHeadSVD.folded_lnHeadSVD.headHeadSVD.is_degenerate()HeadSVD.layerHeadSVD.null_rtolHeadSVD.rank_reportHeadSVD.require_isolated()HeadSVD.which
JSpaceDecompositionJSpaceDecomposition.supportJSpaceDecomposition.coordinatesJSpaceDecomposition.selected_supportJSpaceDecomposition.reconstructionJSpaceDecomposition.j_space_componentJSpaceDecomposition.non_j_space_componentJSpaceDecomposition.coordinatesJSpaceDecomposition.j_space_componentJSpaceDecomposition.non_j_space_componentJSpaceDecomposition.reconstructionJSpaceDecomposition.selected_supportJSpaceDecomposition.support
JSpaceOccupancyJSpaceVarianceProfileJacobianLensJacobianLens.jacobiansJacobianLens.n_promptsJacobianLens.d_modelJacobianLens.metadataJacobianLens.ablation_hooks()JacobianLens.clear_device_cache()JacobianLens.coordinate_patch()JacobianLens.coordinate_patch_hooks()JacobianLens.decompose()JacobianLens.fit()JacobianLens.fraction_of_variance()JacobianLens.from_pretrained()JacobianLens.lens_vector_dictionary()JacobianLens.lens_vectors()JacobianLens.load()JacobianLens.merge()JacobianLens.occupancy()JacobianLens.readout()JacobianLens.save()JacobianLens.source_layersJacobianLens.steering_hooks()JacobianLens.swap_clamp_hooks()JacobianLens.swap_hooks()JacobianLens.transport()JacobianLens.validate_model()
JacobianLensReadoutJacobianLensReadout.lens_topk_valuesJacobianLensReadout.lens_topk_indicesJacobianLensReadout.model_topk_valuesJacobianLensReadout.model_topk_indicesJacobianLensReadout.lens_logitsJacobianLensReadout.model_logitsJacobianLensReadout.tokensJacobianLensReadout.positionsJacobianLensReadout.use_jacobianJacobianLensReadout.lens_logitsJacobianLensReadout.lens_topk_indicesJacobianLensReadout.lens_topk_valuesJacobianLensReadout.model_logitsJacobianLensReadout.model_topk_indicesJacobianLensReadout.model_topk_valuesJacobianLensReadout.positionsJacobianLensReadout.tokensJacobianLensReadout.top_tokens()JacobianLensReadout.use_jacobian
LinearGradientFactorsLogitSignatureNodePatchResultPatchResult.head_svdPatchResult.retainedPatchResult.original_metricPatchResult.patched_metricPatchResult.delta_metricPatchResult.baseline_delta_metricPatchResult.gatedPatchResult.baseline_delta_metricPatchResult.delta_metricPatchResult.gatedPatchResult.head_svdPatchResult.original_metricPatchResult.patched_metricPatchResult.retained
ProjectionKernelResultRandomSubspaceReferenceRankReportRowSparseProbeControlSparseProbeMetricsSparseProbeResultSparseProbeResult.class_weightSparseProbeResult.coefficientsSparseProbeResult.constant_featuresSparseProbeResult.decrement_toleranceSparseProbeResult.feature_scoresSparseProbeResult.function_evaluationsSparseProbeResult.gradient_inf_normSparseProbeResult.interceptSparseProbeResult.iterationsSparseProbeResult.kSparseProbeResult.l2_strengthSparseProbeResult.max_iterSparseProbeResult.max_refinement_stepsSparseProbeResult.metricsSparseProbeResult.negative_labelSparseProbeResult.newton_decrementSparseProbeResult.objectiveSparseProbeResult.positive_labelSparseProbeResult.preprocessSparseProbeResult.preprocess_meanSparseProbeResult.preprocess_scaleSparseProbeResult.refinement_stepsSparseProbeResult.seedSparseProbeResult.selected_featuresSparseProbeResult.stop_reasonSparseProbeResult.test_fractionSparseProbeResult.test_indicesSparseProbeResult.test_negative_countSparseProbeResult.test_positive_countSparseProbeResult.train_indicesSparseProbeResult.train_negative_countSparseProbeResult.train_positive_count
SparseProbeSweepSubspaceBasisSubspaceBasis.basisSubspaceBasis.singular_valuesSubspaceBasis.rankSubspaceBasis.measured_rankSubspaceBasis.rtolSubspaceBasis.thresholdSubspaceBasis.input_shapeSubspaceBasis.ambient_dimSubspaceBasis.basisSubspaceBasis.input_shapeSubspaceBasis.measured_rankSubspaceBasis.rankSubspaceBasis.rtolSubspaceBasis.singular_valuesSubspaceBasis.threshold
VocabularyRankingattention_head_subspace_affinity()attribution_patch()decompose_head()direct_logit_attribution()estimate_occupancy()fit_sparse_probe()get_act_patch_direct_path()get_act_patch_direct_path_all_sources()get_sparse_decomposition()load_artifact()logit_signature()orthonormal_subspace()patch_along_directions()project_activations()projection_kernel()random_projection_kernel_moments()run_causal_swap_benchmark()select_displacement_matched_control_token()solve_coordinate_patch()solve_coordinate_patch_positions()success_rate_ci()sweep_sparse_probe()vocab_readout()
- Submodules
- transformer_lens.tools.model_registry package
- Submodules
- transformer_lens.tools.model_registry.alias_drift module
- transformer_lens.tools.model_registry.api module
- transformer_lens.tools.model_registry.checkpoints module
- transformer_lens.tools.model_registry.discover_architectures module
- transformer_lens.tools.model_registry.exceptions module
- transformer_lens.tools.model_registry.generate_report module
- transformer_lens.tools.model_registry.hf_scraper module
- transformer_lens.tools.model_registry.registry_io module
add_verification_record()extract_phase_scores()is_hf_loadable_quantized()is_incompatible_quantized()is_quantized_model()load_model_aliases()load_supported_models_raw()load_verification_history()pass_status()recompute_registry_totals()registry_prompt_profile()required_quant_library_for_model()resolve_model_alias()save_supported_models_raw()save_verification_history()update_model_status()
- transformer_lens.tools.model_registry.relevancy module
- transformer_lens.tools.model_registry.schemas module
- transformer_lens.tools.model_registry.validate module
- transformer_lens.tools.model_registry.verification module
- transformer_lens.tools.model_registry.verify_models module
- Module contents
ArchitectureAnalysisArchitectureAnalysis.architecture_idArchitectureAnalysis.total_modelsArchitectureAnalysis.total_downloadsArchitectureAnalysis.priority_scoreArchitectureAnalysis.top_modelsArchitectureAnalysis.architecture_idArchitectureAnalysis.priority_scoreArchitectureAnalysis.to_dict()ArchitectureAnalysis.top_modelsArchitectureAnalysis.total_downloadsArchitectureAnalysis.total_models
ArchitectureGapArchitectureGap.architecture_idArchitectureGap.total_modelsArchitectureGap.sample_modelsArchitectureGap.total_downloadsArchitectureGap.min_param_countArchitectureGap.relevancy_scoreArchitectureGap.architecture_idArchitectureGap.from_dict()ArchitectureGap.min_param_countArchitectureGap.relevancy_scoreArchitectureGap.sample_modelsArchitectureGap.to_dict()ArchitectureGap.total_downloadsArchitectureGap.total_models
ArchitectureGapsReportArchitectureGapsReport.generated_atArchitectureGapsReport.scan_infoArchitectureGapsReport.total_unsupported_architecturesArchitectureGapsReport.total_unsupported_modelsArchitectureGapsReport.gapsArchitectureGapsReport.from_dict()ArchitectureGapsReport.gapsArchitectureGapsReport.generated_atArchitectureGapsReport.scan_infoArchitectureGapsReport.to_dict()ArchitectureGapsReport.total_unsupported_architecturesArchitectureGapsReport.total_unsupported_models
ArchitectureNotSupportedErrorArchitectureStatsArchitectureStats.architecture_idArchitectureStats.is_supportedArchitectureStats.model_countArchitectureStats.verified_countArchitectureStats.example_modelsArchitectureStats.architecture_idArchitectureStats.example_modelsArchitectureStats.is_supportedArchitectureStats.model_countArchitectureStats.to_dict()ArchitectureStats.verified_count
DataNotLoadedErrorDataValidationErrorModelEntryModelEntry.architecture_idModelEntry.model_idModelEntry.statusModelEntry.verified_dateModelEntry.metadataModelEntry.noteModelEntry.prompt_profileModelEntry.phase1_scoreModelEntry.phase2_scoreModelEntry.phase3_scoreModelEntry.phase4_scoreModelEntry.phase7_scoreModelEntry.phase8_scoreModelEntry.phase9_scoreModelEntry.architecture_idModelEntry.from_dict()ModelEntry.metadataModelEntry.model_idModelEntry.noteModelEntry.p4_scoring_versionModelEntry.phase1_scoreModelEntry.phase2_scoreModelEntry.phase3_scoreModelEntry.phase4_scoreModelEntry.phase7_scoreModelEntry.phase8_scoreModelEntry.phase9_scoreModelEntry.prompt_profileModelEntry.statusModelEntry.to_dict()ModelEntry.verified_date
ModelMetadataModelNotFoundErrorModelRegistryErrorScanInfoSupportedModelsReportSupportedModelsReport.generated_atSupportedModelsReport.scan_infoSupportedModelsReport.total_architecturesSupportedModelsReport.total_modelsSupportedModelsReport.total_verifiedSupportedModelsReport.modelsSupportedModelsReport.from_dict()SupportedModelsReport.generated_atSupportedModelsReport.modelsSupportedModelsReport.scan_infoSupportedModelsReport.to_dict()SupportedModelsReport.total_architecturesSupportedModelsReport.total_modelsSupportedModelsReport.total_provisionalSupportedModelsReport.total_verified
VerificationHistoryVerificationHistory.recordsVerificationHistory.last_updatedVerificationHistory.add_record()VerificationHistory.from_dict()VerificationHistory.get_record()VerificationHistory.invalidate()VerificationHistory.is_verified()VerificationHistory.last_updatedVerificationHistory.recordsVerificationHistory.to_dict()
VerificationRecordVerificationRecord.model_idVerificationRecord.architecture_idVerificationRecord.verified_dateVerificationRecord.verified_byVerificationRecord.transformerlens_versionVerificationRecord.notesVerificationRecord.invalidatedVerificationRecord.invalidation_reasonVerificationRecord.architecture_idVerificationRecord.from_dict()VerificationRecord.invalidatedVerificationRecord.invalidation_reasonVerificationRecord.model_idVerificationRecord.notesVerificationRecord.p4_scoring_versionVerificationRecord.prompt_profileVerificationRecord.to_dict()VerificationRecord.transformerlens_versionVerificationRecord.verified_byVerificationRecord.verified_date
- Submodules
Submodules¶
- transformer_lens.tools.training module
TrainConfigTrainConfig.batch_sizeTrainConfig.deviceTrainConfig.lrTrainConfig.max_grad_normTrainConfig.max_stepsTrainConfig.momentumTrainConfig.num_epochsTrainConfig.optimizer_nameTrainConfig.print_everyTrainConfig.save_dirTrainConfig.save_everyTrainConfig.seedTrainConfig.wandbTrainConfig.wandb_project_nameTrainConfig.warmup_stepsTrainConfig.weight_decay
train()
Module contents¶
TransformerLens tools package.
This package contains utilities and tools for working with TransformerLens, including the model registry for discovering compatible HuggingFace models.
- Subpackages:
analysis: High-level interpretability analyses (e.g. Direct Logit Attribution)
model_registry: Tools for discovering and documenting supported models
Subpackages load lazily (PEP 562): analysis pulls in torch-heavy core
modules, so an eager import here would create a cycle for core modules that
consume model_registry data at import time.