LICENSE
README.md
pyproject.toml
setup.py
examples/__init__.py
examples/detect_bad_code.py
examples/detect_bias.py
examples/detect_gender_bias.py
examples/detect_hallucinations.py
examples/detect_hallucinations_is_harmful_vs_blocked.py
examples/detect_harmful_content.py
examples/detect_personal_info.py
examples/detect_scheming.py
examples/evaluate_personal.py
examples/just_inference.py
examples/livecodebench_evaluation_example.py
examples/synthetic_steering_optimization.py
examples/test_llama_steering.py
examples/test_multi_property.py
examples/train_classifiers_for_different_models.py
examples/train_different_detectors.py
examples/utils/__init__.py
examples/utils/content_detector.py
examples/utils/load_model.py
wisent_guard/__init__.py
wisent_guard/__main__.py
wisent_guard/cli.py
wisent_guard/guard.py
wisent_guard/harness.py
wisent_guard/inference.py
wisent_guard/optimize.py
wisent_guard/vectors.py
wisent_guard.egg-info/PKG-INFO
wisent_guard.egg-info/SOURCES.txt
wisent_guard.egg-info/dependency_links.txt
wisent_guard.egg-info/requires.txt
wisent_guard.egg-info/top_level.txt
wisent_guard/core/__init__.py
wisent_guard/core/activation_collection_method.py
wisent_guard/core/activations.py
wisent_guard/core/aggregation.py
wisent_guard/core/autonomous_agent.py
wisent_guard/core/benchmark_extractors.py
wisent_guard/core/bigcode_extractors.py
wisent_guard/core/bigcode_integration.py
wisent_guard/core/classification_optimizer.py
wisent_guard/core/classifier.py
wisent_guard/core/control_vector.py
wisent_guard/core/detection_handling.py
wisent_guard/core/ground_truth_evaluator.py
wisent_guard/core/hyperparameter_optimizer.py
wisent_guard/core/layer.py
wisent_guard/core/lm_eval_harness_ground_truth.py
wisent_guard/core/log_likelihoods_evaluator.py
wisent_guard/core/managed_cached_benchmarks.py
wisent_guard/core/mixed_benchmark_sampler.py
wisent_guard/core/model.py
wisent_guard/core/model_config_manager.py
wisent_guard/core/model_persistence.py
wisent_guard/core/normalization.py
wisent_guard/core/parameters.py
wisent_guard/core/parser.py
wisent_guard/core/representation.py
wisent_guard/core/response.py
wisent_guard/core/sample_size_optimizer.py
wisent_guard/core/sample_size_optimizer_v2.py
wisent_guard/core/save_results.py
wisent_guard/core/secure_code_evaluator.py
wisent_guard/core/steering.py
wisent_guard/core/steering_method.py
wisent_guard/core/steering_optimizer.py
wisent_guard/core/task_agnostic_cli.py
wisent_guard/core/task_interface.py
wisent_guard/core/task_selector.py
wisent_guard/core/time_estimator.py
wisent_guard/core/timing_calibration.py
wisent_guard/core/user_model_config.py
wisent_guard/core/agent/__init__.py
wisent_guard/core/agent/budget.py
wisent_guard/core/agent/device_benchmarks.py
wisent_guard/core/agent/diagnose.py
wisent_guard/core/agent/steer.py
wisent_guard/core/agent/timeout.py
wisent_guard/core/agent/diagnose/__init__.py
wisent_guard/core/agent/diagnose/agent_classifier_decision.py
wisent_guard/core/agent/diagnose/classifier_marketplace.py
wisent_guard/core/agent/diagnose/create_classifier.py
wisent_guard/core/agent/diagnose/response_diagnostics.py
wisent_guard/core/agent/diagnose/select_classifiers.py
wisent_guard/core/agent/diagnose/synthetic_classifier_option.py
wisent_guard/core/agent/diagnose/test_synthetic_classifier.py
wisent_guard/core/agent/diagnose/tasks/__init__.py
wisent_guard/core/agent/diagnose/tasks/task_manager.py
wisent_guard/core/agent/diagnose/tasks/task_relevance.py
wisent_guard/core/agent/diagnose/tasks/task_selector.py
wisent_guard/core/benchmark_extractor_impls/__init__.py
wisent_guard/core/benchmark_extractor_impls/base.py
wisent_guard/core/benchmark_extractor_impls/livecodebench_model_outputs_extractor.py
wisent_guard/core/contrastive_pairs/__init__.py
wisent_guard/core/contrastive_pairs/contrastive_database.py
wisent_guard/core/contrastive_pairs/contrastive_generation_conf.py
wisent_guard/core/contrastive_pairs/contrastive_pair.py
wisent_guard/core/contrastive_pairs/contrastive_pair_set.py
wisent_guard/core/contrastive_pairs/generate_synthetically.py
wisent_guard/core/contrastive_pairs/quality_check.py
wisent_guard/core/contrastive_pairs/question_bank.py
wisent_guard/core/contrastive_pairs/question_bank_cli.py
wisent_guard/core/data_loaders/__init__.py
wisent_guard/core/data_loaders/livecodebench_loader.py
wisent_guard/core/data_loaders/livecodebench_model_outputs.py
wisent_guard/core/data_loaders/steering_data_extractor.py
wisent_guard/core/data_loaders/version_manager.py
wisent_guard/core/docker/__init__.py
wisent_guard/core/docker/optimized_docker_executor.py
wisent_guard/core/evaluate/__init__.py
wisent_guard/core/evaluate/stop_nonsense.py
wisent_guard/core/evaluation/__init__.py
wisent_guard/core/evaluation/benchmarks.py
wisent_guard/core/evaluation/exporters.py
wisent_guard/core/models/__init__.py
wisent_guard/core/models/steering_compatible_model.py
wisent_guard/core/optuna/__init__.py
wisent_guard/core/optuna/data_utils.py
wisent_guard/core/optuna/metrics.py
wisent_guard/core/optuna/optuna_pipeline.py
wisent_guard/core/optuna/pipeline.py
wisent_guard/core/optuna/probe_optimization.py
wisent_guard/core/optuna/steering_optimization.py
wisent_guard/core/pipelines/__init__.py
wisent_guard/core/pipelines/activation_collector.py
wisent_guard/core/pipelines/experiment_runner.py
wisent_guard/core/pipelines/steering_trainer.py
wisent_guard/core/steering_methods/__init__.py
wisent_guard/core/steering_methods/base.py
wisent_guard/core/steering_methods/bipo.py
wisent_guard/core/steering_methods/caa.py
wisent_guard/core/steering_methods/control_vector_steering.py
wisent_guard/core/steering_methods/dac.py
wisent_guard/core/steering_methods/hpr.py
wisent_guard/core/steering_methods/k_steering.py
wisent_guard/core/steering_methods/steering_evaluation.py
wisent_guard/core/steering_methods/token_steered.py
wisent_guard/core/tasks/__init__.py
wisent_guard/core/tasks/aime_task.py
wisent_guard/core/tasks/file_task.py
wisent_guard/core/tasks/hle_task.py
wisent_guard/core/tasks/hmmt_task.py
wisent_guard/core/tasks/livecodebench_task.py
wisent_guard/core/tasks/livemathbench_task.py
wisent_guard/core/tasks/lm_eval_task.py
wisent_guard/core/tasks/math500_task.py
wisent_guard/core/tasks/polymath_task.py
wisent_guard/core/tasks/supergpqa_task.py
wisent_guard/core/tracking/__init__.py
wisent_guard/core/tracking/latency.py
wisent_guard/core/tracking/memory.py
wisent_guard/utils/__init__.py
wisent_guard/utils/evaluate_layer_performance.py
wisent_guard/utils/evaluate_layer_performance_truthfulqa.py
wisent_guard/utils/helpers.py
wisent_guard/utils/logger.py