| { |
| "manifest_schema_version": 3, |
| "config_path": "/nvme/CognitiveMemoryLayer/packages/models/model_pipeline.toml", |
| "trained_at_utc": "2026-03-25T13:59:08.669850+00:00", |
| "paths": { |
| "prepared_dir": "/nvme/CognitiveMemoryLayer/packages/models/prepared_data/modelpack", |
| "trained_models_dir": "/nvme/CognitiveMemoryLayer/packages/models/trained_models" |
| }, |
| "train_settings": { |
| "seed": 42, |
| "families": [ |
| "router", |
| "extractor", |
| "pair" |
| ], |
| "max_features": 250000, |
| "min_df": 2, |
| "ngram_min": 1, |
| "ngram_max": 2, |
| "max_iter": 25, |
| "alpha": 1e-05, |
| "predict_batch_size": 8192, |
| "early_stopping": true, |
| "early_stopping_patience": 3, |
| "early_stopping_metric": "macro_f1", |
| "early_stopping_min_delta": 0.001, |
| "calibration_method": "sigmoid", |
| "calibration_split": "eval", |
| "transformer": { |
| "model_name_or_path": "microsoft/deberta-v3-base", |
| "tokenizer_name": "", |
| "num_train_epochs": 2, |
| "per_device_train_batch_size": 8, |
| "per_device_eval_batch_size": 16, |
| "max_seq_length": 256, |
| "learning_rate": 2e-05, |
| "warmup_ratio": 0.1, |
| "weight_decay": 0.01, |
| "gradient_accumulation_steps": 1, |
| "score_margin": 0.15, |
| "focal_gamma": 1.5, |
| "temperature_grid": [ |
| 0.7, |
| 0.85, |
| 1.0, |
| 1.15, |
| 1.3, |
| 1.5, |
| 2.0 |
| ] |
| }, |
| "token": { |
| "model_name_or_path": "bert-base-multilingual-cased", |
| "num_train_epochs": 8, |
| "per_device_train_batch_size": 8, |
| "per_device_eval_batch_size": 16, |
| "max_seq_length": 256, |
| "stride": 64, |
| "learning_rate": 5e-05, |
| "warmup_ratio": 0.1, |
| "weight_decay": 0.01, |
| "gradient_accumulation_steps": 1 |
| }, |
| "allow_dirty": false, |
| "release_mode": false, |
| "strict": true |
| }, |
| "build_metadata": { |
| "python_version": "3.14.3 (main, Mar 11 2026, 19:17:30) [GCC 15.2.0]", |
| "dependencies": { |
| "scikit_learn": "1.8.0", |
| "joblib": "1.5.3", |
| "pandas": "3.0.1" |
| }, |
| "commit_sha": "090060c4478c373a02eb09ae63292f51ca31a710", |
| "dirty": true |
| }, |
| "configured_tasks": [ |
| { |
| "task_name": "retrieval_constraint_relevance_pair", |
| "family": "pair", |
| "input_type": "pair", |
| "objective": "pair_ranking", |
| "enabled": true, |
| "artifact_name": "retrieval_constraint_relevance_pair", |
| "metrics": [ |
| "mrr@10", |
| "ndcg@10", |
| "recall@10" |
| ], |
| "trainer": "transformer_pair", |
| "feature_backend": "", |
| "label_order": [], |
| "embedding_model_name": "", |
| "backbone_model_name": "BAAI/bge-reranker-base", |
| "tokenizer_name": "" |
| }, |
| { |
| "task_name": "memory_rerank_pair", |
| "family": "pair", |
| "input_type": "pair", |
| "objective": "pair_ranking", |
| "enabled": true, |
| "artifact_name": "memory_rerank_pair", |
| "metrics": [ |
| "mrr@10", |
| "ndcg@10", |
| "recall@10" |
| ], |
| "trainer": "transformer_pair", |
| "feature_backend": "", |
| "label_order": [], |
| "embedding_model_name": "", |
| "backbone_model_name": "BAAI/bge-reranker-base", |
| "tokenizer_name": "" |
| }, |
| { |
| "task_name": "novelty_pair", |
| "family": "pair", |
| "input_type": "pair", |
| "objective": "classification", |
| "enabled": true, |
| "artifact_name": "novelty_pair", |
| "metrics": [ |
| "accuracy", |
| "macro_f1", |
| "weighted_f1" |
| ], |
| "trainer": "transformer_pair", |
| "feature_backend": "", |
| "label_order": [], |
| "embedding_model_name": "", |
| "backbone_model_name": "BAAI/bge-reranker-base", |
| "tokenizer_name": "" |
| }, |
| { |
| "task_name": "fact_extraction_structured", |
| "family": "extractor", |
| "input_type": "single", |
| "objective": "token_classification", |
| "enabled": true, |
| "artifact_name": "fact_extraction_structured", |
| "metrics": [ |
| "span_f1", |
| "span_exact_match" |
| ], |
| "trainer": "", |
| "feature_backend": "", |
| "label_order": [], |
| "embedding_model_name": "", |
| "backbone_model_name": "", |
| "tokenizer_name": "" |
| }, |
| { |
| "task_name": "schema_match_pair", |
| "family": "pair", |
| "input_type": "pair", |
| "objective": "classification", |
| "enabled": true, |
| "artifact_name": "schema_match_pair", |
| "metrics": [ |
| "accuracy", |
| "macro_f1", |
| "weighted_f1" |
| ], |
| "trainer": "transformer_pair", |
| "feature_backend": "", |
| "label_order": [], |
| "embedding_model_name": "", |
| "backbone_model_name": "BAAI/bge-reranker-base", |
| "tokenizer_name": "" |
| }, |
| { |
| "task_name": "reconsolidation_candidate_pair", |
| "family": "pair", |
| "input_type": "pair", |
| "objective": "pair_ranking", |
| "enabled": true, |
| "artifact_name": "reconsolidation_candidate_pair", |
| "metrics": [ |
| "mrr@10", |
| "ndcg@10", |
| "recall@10" |
| ], |
| "trainer": "transformer_pair", |
| "feature_backend": "", |
| "label_order": [], |
| "embedding_model_name": "", |
| "backbone_model_name": "BAAI/bge-reranker-base", |
| "tokenizer_name": "" |
| }, |
| { |
| "task_name": "write_importance_regression", |
| "family": "router", |
| "input_type": "single", |
| "objective": "single_regression", |
| "enabled": true, |
| "artifact_name": "write_importance_regression", |
| "metrics": [ |
| "mae", |
| "rmse" |
| ], |
| "trainer": "", |
| "feature_backend": "", |
| "label_order": [], |
| "embedding_model_name": "", |
| "backbone_model_name": "", |
| "tokenizer_name": "" |
| }, |
| { |
| "task_name": "memory_type", |
| "family": "router", |
| "input_type": "single", |
| "objective": "classification", |
| "enabled": true, |
| "artifact_name": "memory_type", |
| "metrics": [ |
| "accuracy", |
| "macro_f1", |
| "weighted_f1" |
| ], |
| "trainer": "hierarchical_transformer", |
| "feature_backend": "", |
| "label_order": [], |
| "embedding_model_name": "", |
| "backbone_model_name": "microsoft/deberta-v3-base", |
| "tokenizer_name": "" |
| }, |
| { |
| "task_name": "salience_bin", |
| "family": "router", |
| "input_type": "single", |
| "objective": "classification", |
| "enabled": true, |
| "artifact_name": "salience_bin", |
| "metrics": [ |
| "accuracy", |
| "macro_f1", |
| "weighted_f1", |
| "ordinal_mae", |
| "off_by_two_rate" |
| ], |
| "trainer": "ordinal_threshold", |
| "feature_backend": "", |
| "label_order": [ |
| "low", |
| "medium", |
| "high" |
| ], |
| "embedding_model_name": "", |
| "backbone_model_name": "", |
| "tokenizer_name": "" |
| }, |
| { |
| "task_name": "importance_bin", |
| "family": "router", |
| "input_type": "single", |
| "objective": "classification", |
| "enabled": true, |
| "artifact_name": "importance_bin", |
| "metrics": [ |
| "accuracy", |
| "macro_f1", |
| "weighted_f1", |
| "ordinal_mae", |
| "off_by_two_rate" |
| ], |
| "trainer": "ordinal_threshold", |
| "feature_backend": "", |
| "label_order": [ |
| "low", |
| "medium", |
| "high" |
| ], |
| "embedding_model_name": "", |
| "backbone_model_name": "", |
| "tokenizer_name": "" |
| }, |
| { |
| "task_name": "confidence_bin", |
| "family": "router", |
| "input_type": "single", |
| "objective": "classification", |
| "enabled": true, |
| "artifact_name": "confidence_bin", |
| "metrics": [ |
| "accuracy", |
| "macro_f1", |
| "weighted_f1", |
| "ordinal_mae", |
| "off_by_two_rate" |
| ], |
| "trainer": "ordinal_threshold", |
| "feature_backend": "", |
| "label_order": [ |
| "low", |
| "medium", |
| "high" |
| ], |
| "embedding_model_name": "", |
| "backbone_model_name": "", |
| "tokenizer_name": "" |
| }, |
| { |
| "task_name": "decay_profile", |
| "family": "router", |
| "input_type": "single", |
| "objective": "classification", |
| "enabled": true, |
| "artifact_name": "decay_profile", |
| "metrics": [ |
| "accuracy", |
| "macro_f1", |
| "weighted_f1", |
| "ordinal_mae", |
| "off_by_two_rate" |
| ], |
| "trainer": "ordinal_threshold", |
| "feature_backend": "", |
| "label_order": [ |
| "very_fast", |
| "fast", |
| "medium", |
| "slow", |
| "very_slow" |
| ], |
| "embedding_model_name": "", |
| "backbone_model_name": "", |
| "tokenizer_name": "" |
| }, |
| { |
| "task_name": "pii_span_detection", |
| "family": "extractor", |
| "input_type": "single", |
| "objective": "token_classification", |
| "enabled": true, |
| "artifact_name": "pii_span_detection", |
| "metrics": [ |
| "span_f1", |
| "span_exact_match" |
| ], |
| "trainer": "", |
| "feature_backend": "", |
| "label_order": [], |
| "embedding_model_name": "", |
| "backbone_model_name": "", |
| "tokenizer_name": "" |
| }, |
| { |
| "task_name": "consolidation_gist_quality", |
| "family": "router", |
| "input_type": "single", |
| "objective": "classification", |
| "enabled": true, |
| "artifact_name": "consolidation_gist_quality", |
| "metrics": [ |
| "accuracy", |
| "macro_f1", |
| "weighted_f1" |
| ], |
| "trainer": "transformer_text", |
| "feature_backend": "", |
| "label_order": [], |
| "embedding_model_name": "", |
| "backbone_model_name": "microsoft/deberta-v3-base", |
| "tokenizer_name": "" |
| }, |
| { |
| "task_name": "forgetting_action_policy", |
| "family": "router", |
| "input_type": "single", |
| "objective": "classification", |
| "enabled": true, |
| "artifact_name": "forgetting_action_policy", |
| "metrics": [ |
| "accuracy", |
| "macro_f1", |
| "weighted_f1" |
| ], |
| "trainer": "transformer_text", |
| "feature_backend": "", |
| "label_order": [], |
| "embedding_model_name": "", |
| "backbone_model_name": "microsoft/deberta-v3-base", |
| "tokenizer_name": "" |
| }, |
| { |
| "task_name": "constraint_dimension", |
| "family": "router", |
| "input_type": "single", |
| "objective": "classification", |
| "enabled": true, |
| "artifact_name": "constraint_dimension", |
| "metrics": [ |
| "accuracy", |
| "macro_f1", |
| "weighted_f1" |
| ], |
| "trainer": "transformer_text", |
| "feature_backend": "", |
| "label_order": [], |
| "embedding_model_name": "", |
| "backbone_model_name": "microsoft/deberta-v3-base", |
| "tokenizer_name": "" |
| }, |
| { |
| "task_name": "context_tag", |
| "family": "router", |
| "input_type": "single", |
| "objective": "classification", |
| "enabled": true, |
| "artifact_name": "context_tag", |
| "metrics": [ |
| "accuracy", |
| "macro_f1", |
| "weighted_f1" |
| ], |
| "trainer": "transformer_text", |
| "feature_backend": "", |
| "label_order": [], |
| "embedding_model_name": "", |
| "backbone_model_name": "microsoft/deberta-v3-base", |
| "tokenizer_name": "" |
| } |
| ], |
| "preflight_validation": { |
| "ok": true, |
| "strict": true, |
| "errors": [], |
| "warnings": [], |
| "task_checks": [ |
| { |
| "task_name": "retrieval_constraint_relevance_pair", |
| "family": "pair", |
| "input_type": "pair", |
| "objective": "pair_ranking", |
| "enabled": true, |
| "status": "ok", |
| "reason": null, |
| "rows_found": 80000, |
| "valid_score_rows": 0 |
| }, |
| { |
| "task_name": "memory_rerank_pair", |
| "family": "pair", |
| "input_type": "pair", |
| "objective": "pair_ranking", |
| "enabled": true, |
| "status": "ok", |
| "reason": null, |
| "rows_found": 80000, |
| "valid_score_rows": 0 |
| }, |
| { |
| "task_name": "novelty_pair", |
| "family": "pair", |
| "input_type": "pair", |
| "objective": "classification", |
| "enabled": true, |
| "status": "ok", |
| "reason": null, |
| "rows_found": 117206, |
| "valid_score_rows": 0 |
| }, |
| { |
| "task_name": "fact_extraction_structured", |
| "family": "extractor", |
| "input_type": "single", |
| "objective": "token_classification", |
| "enabled": true, |
| "status": "ok", |
| "reason": null, |
| "rows_found": 31997, |
| "valid_score_rows": 0 |
| }, |
| { |
| "task_name": "schema_match_pair", |
| "family": "pair", |
| "input_type": "pair", |
| "objective": "classification", |
| "enabled": true, |
| "status": "ok", |
| "reason": null, |
| "rows_found": 81200, |
| "valid_score_rows": 0 |
| }, |
| { |
| "task_name": "reconsolidation_candidate_pair", |
| "family": "pair", |
| "input_type": "pair", |
| "objective": "pair_ranking", |
| "enabled": true, |
| "status": "ok", |
| "reason": null, |
| "rows_found": 80000, |
| "valid_score_rows": 0 |
| }, |
| { |
| "task_name": "write_importance_regression", |
| "family": "router", |
| "input_type": "single", |
| "objective": "single_regression", |
| "enabled": true, |
| "status": "ok", |
| "reason": null, |
| "rows_found": 8000, |
| "valid_score_rows": 8000 |
| }, |
| { |
| "task_name": "memory_type", |
| "family": "router", |
| "input_type": "single", |
| "objective": "classification", |
| "enabled": true, |
| "status": "ok", |
| "reason": null, |
| "rows_found": 120000, |
| "valid_score_rows": 0 |
| }, |
| { |
| "task_name": "salience_bin", |
| "family": "router", |
| "input_type": "single", |
| "objective": "classification", |
| "enabled": true, |
| "status": "ok", |
| "reason": null, |
| "rows_found": 24000, |
| "valid_score_rows": 0 |
| }, |
| { |
| "task_name": "importance_bin", |
| "family": "router", |
| "input_type": "single", |
| "objective": "classification", |
| "enabled": true, |
| "status": "ok", |
| "reason": null, |
| "rows_found": 24000, |
| "valid_score_rows": 0 |
| }, |
| { |
| "task_name": "confidence_bin", |
| "family": "router", |
| "input_type": "single", |
| "objective": "classification", |
| "enabled": true, |
| "status": "ok", |
| "reason": null, |
| "rows_found": 24000, |
| "valid_score_rows": 0 |
| }, |
| { |
| "task_name": "decay_profile", |
| "family": "router", |
| "input_type": "single", |
| "objective": "classification", |
| "enabled": true, |
| "status": "ok", |
| "reason": null, |
| "rows_found": 40000, |
| "valid_score_rows": 0 |
| }, |
| { |
| "task_name": "pii_span_detection", |
| "family": "extractor", |
| "input_type": "single", |
| "objective": "token_classification", |
| "enabled": true, |
| "status": "ok", |
| "reason": null, |
| "rows_found": 30516, |
| "valid_score_rows": 0 |
| }, |
| { |
| "task_name": "consolidation_gist_quality", |
| "family": "router", |
| "input_type": "single", |
| "objective": "classification", |
| "enabled": true, |
| "status": "ok", |
| "reason": null, |
| "rows_found": 10432, |
| "valid_score_rows": 0 |
| }, |
| { |
| "task_name": "forgetting_action_policy", |
| "family": "router", |
| "input_type": "single", |
| "objective": "classification", |
| "enabled": true, |
| "status": "ok", |
| "reason": null, |
| "rows_found": 40000, |
| "valid_score_rows": 0 |
| }, |
| { |
| "task_name": "constraint_dimension", |
| "family": "router", |
| "input_type": "single", |
| "objective": "classification", |
| "enabled": true, |
| "status": "ok", |
| "reason": null, |
| "rows_found": 48000, |
| "valid_score_rows": 0 |
| }, |
| { |
| "task_name": "context_tag", |
| "family": "router", |
| "input_type": "single", |
| "objective": "classification", |
| "enabled": true, |
| "status": "ok", |
| "reason": null, |
| "rows_found": 64000, |
| "valid_score_rows": 0 |
| } |
| ], |
| "observed_tasks_by_family": { |
| "extractor": [ |
| "constraint_scope", |
| "constraint_stability", |
| "constraint_type", |
| "fact_type", |
| "pii_presence" |
| ], |
| "pair": [ |
| "conflict_detection", |
| "constraint_rerank", |
| "memory_rerank_pair", |
| "novelty_pair", |
| "reconsolidation_candidate_pair", |
| "retrieval_constraint_relevance_pair", |
| "schema_match_pair", |
| "scope_match", |
| "supersession" |
| ], |
| "router": [ |
| "confidence_bin", |
| "consolidation_gist_quality", |
| "constraint_dimension", |
| "context_tag", |
| "decay_profile", |
| "forgetting_action_policy", |
| "importance_bin", |
| "memory_type", |
| "query_domain", |
| "query_intent", |
| "salience_bin", |
| "write_importance_regression" |
| ] |
| }, |
| "coverage_vs_config": { |
| "extractor": { |
| "configured_enabled_tasks": [], |
| "observed_tasks": [ |
| "constraint_scope", |
| "constraint_stability", |
| "constraint_type", |
| "fact_type", |
| "pii_presence" |
| ], |
| "missing_configured_tasks": [] |
| }, |
| "pair": { |
| "configured_enabled_tasks": [ |
| "memory_rerank_pair", |
| "novelty_pair", |
| "reconsolidation_candidate_pair", |
| "retrieval_constraint_relevance_pair", |
| "schema_match_pair" |
| ], |
| "observed_tasks": [ |
| "conflict_detection", |
| "constraint_rerank", |
| "memory_rerank_pair", |
| "novelty_pair", |
| "reconsolidation_candidate_pair", |
| "retrieval_constraint_relevance_pair", |
| "schema_match_pair", |
| "scope_match", |
| "supersession" |
| ], |
| "missing_configured_tasks": [] |
| }, |
| "router": { |
| "configured_enabled_tasks": [ |
| "confidence_bin", |
| "consolidation_gist_quality", |
| "constraint_dimension", |
| "context_tag", |
| "decay_profile", |
| "forgetting_action_policy", |
| "importance_bin", |
| "memory_type", |
| "salience_bin", |
| "write_importance_regression" |
| ], |
| "observed_tasks": [ |
| "confidence_bin", |
| "consolidation_gist_quality", |
| "constraint_dimension", |
| "context_tag", |
| "decay_profile", |
| "forgetting_action_policy", |
| "importance_bin", |
| "memory_type", |
| "query_domain", |
| "query_intent", |
| "salience_bin", |
| "write_importance_regression" |
| ], |
| "missing_configured_tasks": [] |
| } |
| } |
| }, |
| "families": { |
| "router": { |
| "model_path": "packages/models/trained_models/router_model.joblib" |
| }, |
| "extractor": { |
| "model_path": "packages/models/trained_models/extractor_model.joblib" |
| }, |
| "pair": { |
| "model_path": "packages/models/trained_models/pair_model.joblib" |
| } |
| }, |
| "task_training_status": { |
| "retrieval_constraint_relevance_pair": { |
| "status": "filtered_out", |
| "reason": "Excluded by --tasks filter", |
| "family": "pair", |
| "objective": "pair_ranking", |
| "enabled": true |
| }, |
| "memory_rerank_pair": { |
| "status": "filtered_out", |
| "reason": "Excluded by --tasks filter", |
| "family": "pair", |
| "objective": "pair_ranking", |
| "enabled": true |
| }, |
| "novelty_pair": { |
| "status": "filtered_out", |
| "reason": "Excluded by --tasks filter", |
| "family": "pair", |
| "objective": "classification", |
| "enabled": true |
| }, |
| "fact_extraction_structured": { |
| "status": "filtered_out", |
| "reason": "Excluded by --tasks filter", |
| "family": "extractor", |
| "objective": "token_classification", |
| "enabled": true |
| }, |
| "schema_match_pair": { |
| "status": "filtered_out", |
| "reason": "Excluded by --tasks filter", |
| "family": "pair", |
| "objective": "classification", |
| "enabled": true |
| }, |
| "reconsolidation_candidate_pair": { |
| "status": "filtered_out", |
| "reason": "Excluded by --tasks filter", |
| "family": "pair", |
| "objective": "pair_ranking", |
| "enabled": true |
| }, |
| "write_importance_regression": { |
| "status": "filtered_out", |
| "reason": "Excluded by --tasks filter", |
| "family": "router", |
| "objective": "single_regression", |
| "enabled": true |
| }, |
| "memory_type": { |
| "status": "filtered_out", |
| "reason": "Excluded by --tasks filter", |
| "family": "router", |
| "objective": "classification", |
| "enabled": true |
| }, |
| "salience_bin": { |
| "status": "filtered_out", |
| "reason": "Excluded by --tasks filter", |
| "family": "router", |
| "objective": "classification", |
| "enabled": true |
| }, |
| "importance_bin": { |
| "status": "filtered_out", |
| "reason": "Excluded by --tasks filter", |
| "family": "router", |
| "objective": "classification", |
| "enabled": true |
| }, |
| "confidence_bin": { |
| "status": "filtered_out", |
| "reason": "Excluded by --tasks filter", |
| "family": "router", |
| "objective": "classification", |
| "enabled": true |
| }, |
| "decay_profile": { |
| "status": "filtered_out", |
| "reason": "Excluded by --tasks filter", |
| "family": "router", |
| "objective": "classification", |
| "enabled": true |
| }, |
| "pii_span_detection": { |
| "status": "trained", |
| "reason": null, |
| "family": "extractor", |
| "objective": "token_classification", |
| "enabled": true, |
| "model_path": "/nvme/CognitiveMemoryLayer/packages/models/trained_models/pii_span_detection_model.joblib", |
| "train_rows": 30516 |
| }, |
| "consolidation_gist_quality": { |
| "status": "filtered_out", |
| "reason": "Excluded by --tasks filter", |
| "family": "router", |
| "objective": "classification", |
| "enabled": true |
| }, |
| "forgetting_action_policy": { |
| "status": "filtered_out", |
| "reason": "Excluded by --tasks filter", |
| "family": "router", |
| "objective": "classification", |
| "enabled": true |
| }, |
| "constraint_dimension": { |
| "status": "filtered_out", |
| "reason": "Excluded by --tasks filter", |
| "family": "router", |
| "objective": "classification", |
| "enabled": true |
| }, |
| "context_tag": { |
| "status": "filtered_out", |
| "reason": "Excluded by --tasks filter", |
| "family": "router", |
| "objective": "classification", |
| "enabled": true |
| } |
| }, |
| "task_models": { |
| "memory_rerank_pair": { |
| "model_path": "packages/models/trained_models/memory_rerank_pair_model.joblib", |
| "hf_model_dir": "packages/models/trained_models/memory_rerank_pair_hf" |
| }, |
| "retrieval_constraint_relevance_pair": { |
| "model_path": "packages/models/trained_models/retrieval_constraint_relevance_pair_model.joblib", |
| "hf_model_dir": "packages/models/trained_models/retrieval_constraint_relevance_pair_hf" |
| }, |
| "novelty_pair": { |
| "model_path": "packages/models/trained_models/novelty_pair_model.joblib", |
| "hf_model_dir": "packages/models/trained_models/novelty_pair_hf" |
| }, |
| "fact_extraction_structured": { |
| "model_path": "packages/models/trained_models/fact_extraction_structured_model.joblib", |
| "hf_model_dir": "packages/models/trained_models/fact_extraction_structured_hf" |
| }, |
| "schema_match_pair": { |
| "task": "schema_match_pair", |
| "objective": "classification", |
| "trainer": "transformer_pair", |
| "model_kind": "transformer_pair", |
| "model_path": "/nvme/CognitiveMemoryLayer/packages/models/trained_models/schema_match_pair_model.joblib", |
| "hf_model_dir": "/nvme/CognitiveMemoryLayer/packages/models/trained_models/schema_match_pair_hf", |
| "train_rows": 81200, |
| "test": { |
| "rows": 10000, |
| "accuracy": 0.8552, |
| "macro_f1": 0.8551561470750885, |
| "weighted_f1": 0.8551561470750885, |
| "micro_f1": 0.8552, |
| "labels": [ |
| "schema_match_pair::match", |
| "schema_match_pair::no_match" |
| ], |
| "confusion_matrix": [ |
| [ |
| 4189, |
| 811 |
| ], |
| [ |
| 637, |
| 4363 |
| ] |
| ], |
| "calibration_error": 0.03336120770113563, |
| "classification_report": { |
| "schema_match_pair::match": { |
| "precision": 0.8680066307501036, |
| "recall": 0.8378, |
| "f1-score": 0.852635864034195, |
| "support": 5000.0 |
| }, |
| "schema_match_pair::no_match": { |
| "precision": 0.8432547352145342, |
| "recall": 0.8726, |
| "f1-score": 0.8576764301159819, |
| "support": 5000.0 |
| }, |
| "accuracy": 0.8552, |
| "macro avg": { |
| "precision": 0.8556306829823189, |
| "recall": 0.8552, |
| "f1-score": 0.8551561470750885, |
| "support": 10000.0 |
| }, |
| "weighted avg": { |
| "precision": 0.8556306829823188, |
| "recall": 0.8552, |
| "f1-score": 0.8551561470750885, |
| "support": 10000.0 |
| } |
| } |
| }, |
| "eval": { |
| "rows": 10000, |
| "accuracy": 0.8517, |
| "macro_f1": 0.8516351927685533, |
| "weighted_f1": 0.8516351927685533, |
| "micro_f1": 0.8517, |
| "labels": [ |
| "schema_match_pair::match", |
| "schema_match_pair::no_match" |
| ], |
| "confusion_matrix": [ |
| [ |
| 4154, |
| 846 |
| ], |
| [ |
| 637, |
| 4363 |
| ] |
| ], |
| "calibration_error": 0.03950018305884612, |
| "classification_report": { |
| "schema_match_pair::match": { |
| "precision": 0.8670423711125026, |
| "recall": 0.8308, |
| "f1-score": 0.848534368297416, |
| "support": 5000.0 |
| }, |
| "schema_match_pair::no_match": { |
| "precision": 0.8375887886350547, |
| "recall": 0.8726, |
| "f1-score": 0.8547360172396905, |
| "support": 5000.0 |
| }, |
| "accuracy": 0.8517, |
| "macro avg": { |
| "precision": 0.8523155798737787, |
| "recall": 0.8517, |
| "f1-score": 0.8516351927685533, |
| "support": 10000.0 |
| }, |
| "weighted avg": { |
| "precision": 0.8523155798737786, |
| "recall": 0.8517, |
| "f1-score": 0.8516351927685533, |
| "support": 10000.0 |
| } |
| } |
| }, |
| "actual_epochs": 4, |
| "best_epoch": 4, |
| "early_stopped": false, |
| "selection_metric": "macro_f1", |
| "selection_value": 0.8516351927685533, |
| "backbone_model_name": "BAAI/bge-reranker-base", |
| "tokenizer_name": "BAAI/bge-reranker-base", |
| "calibration": { |
| "method": "temperature_grid_search", |
| "rows": 10000, |
| "temperature": 1.3, |
| "loss": 0.3754132442613608 |
| }, |
| "thresholds": { |
| "default_threshold": 0.26, |
| "positive_label": "schema_match_pair::match", |
| "precision_floor": 0.85, |
| "positive_f1": 0.8544423440453687, |
| "positive_precision": 0.8501286873886359 |
| }, |
| "artifact_scope": "task", |
| "evaluation_suite": "standard", |
| "dataset_hashes": { |
| "train": { |
| "path": "/nvme/CognitiveMemoryLayer/packages/models/prepared_data/modelpack/pair_train.parquet", |
| "sha256": "5b19ae2af42d3bbeb58fa8cda2a6b40748e2561a7c4c60a464e2eced711d7c16", |
| "bytes": 99474783 |
| }, |
| "test": { |
| "path": "/nvme/CognitiveMemoryLayer/packages/models/prepared_data/modelpack/pair_test.parquet", |
| "sha256": "33f2d21d29ab684df5c9ab8df2b8bea8ffaa485434bc14bed350cea86a075a88", |
| "bytes": 14270999 |
| }, |
| "eval": { |
| "path": "/nvme/CognitiveMemoryLayer/packages/models/prepared_data/modelpack/pair_eval.parquet", |
| "sha256": "cb6fae7ab447adb0d49af2de458dfbfa82f1af1efa5fd72644e2e6a9db7d07b7", |
| "bytes": 14298411 |
| } |
| }, |
| "release_gates": { |
| "passed": true, |
| "checks": [ |
| { |
| "section": "test", |
| "metric": "macro_f1", |
| "threshold": 0.8, |
| "actual": 0.8551561470750885, |
| "passed": true |
| }, |
| { |
| "section": "test", |
| "metric": "calibration_error", |
| "threshold": { |
| "max": 0.08 |
| }, |
| "actual": 0.03336120770113563, |
| "passed": true |
| } |
| ] |
| } |
| }, |
| "reconsolidation_candidate_pair": { |
| "model_path": "packages/models/trained_models/reconsolidation_candidate_pair_model.joblib", |
| "hf_model_dir": "packages/models/trained_models/reconsolidation_candidate_pair_hf" |
| }, |
| "write_importance_regression": { |
| "model_path": "packages/models/trained_models/write_importance_regression_model.joblib" |
| }, |
| "forgetting_action_policy": { |
| "task": "forgetting_action_policy", |
| "objective": "classification", |
| "trainer": "transformer_text", |
| "model_kind": "transformer_text", |
| "model_path": "/nvme/CognitiveMemoryLayer/packages/models/trained_models/forgetting_action_policy_model.joblib", |
| "hf_model_dir": "/nvme/CognitiveMemoryLayer/packages/models/trained_models/forgetting_action_policy_hf", |
| "train_rows": 28580, |
| "test": { |
| "rows": 18920, |
| "accuracy": 0.9994714587737844, |
| "macro_f1": 0.9994714168187103, |
| "weighted_f1": 0.9994714168187103, |
| "micro_f1": 0.9994714587737844, |
| "labels": [ |
| "forgetting_action_policy::compress", |
| "forgetting_action_policy::decay", |
| "forgetting_action_policy::delete", |
| "forgetting_action_policy::keep", |
| "forgetting_action_policy::silence" |
| ], |
| "confusion_matrix": [ |
| [ |
| 3782, |
| 2, |
| 0, |
| 0, |
| 0 |
| ], |
| [ |
| 0, |
| 3784, |
| 0, |
| 0, |
| 0 |
| ], |
| [ |
| 0, |
| 0, |
| 3779, |
| 1, |
| 4 |
| ], |
| [ |
| 0, |
| 0, |
| 0, |
| 3784, |
| 0 |
| ], |
| [ |
| 2, |
| 0, |
| 1, |
| 0, |
| 3781 |
| ] |
| ], |
| "calibration_error": 0.01649172677837832, |
| "classification_report": { |
| "forgetting_action_policy::compress": { |
| "precision": 0.9994714587737844, |
| "recall": 0.9994714587737844, |
| "f1-score": 0.9994714587737844, |
| "support": 3784.0 |
| }, |
| "forgetting_action_policy::decay": { |
| "precision": 0.9994717379820391, |
| "recall": 1.0, |
| "f1-score": 0.9997357992073976, |
| "support": 3784.0 |
| }, |
| "forgetting_action_policy::delete": { |
| "precision": 0.9997354497354497, |
| "recall": 0.9986786469344608, |
| "f1-score": 0.9992067689053411, |
| "support": 3784.0 |
| }, |
| "forgetting_action_policy::keep": { |
| "precision": 0.9997357992073976, |
| "recall": 1.0, |
| "f1-score": 0.9998678821508786, |
| "support": 3784.0 |
| }, |
| "forgetting_action_policy::silence": { |
| "precision": 0.9989431968295905, |
| "recall": 0.9992071881606766, |
| "f1-score": 0.9990751750561501, |
| "support": 3784.0 |
| }, |
| "accuracy": 0.9994714587737844, |
| "macro avg": { |
| "precision": 0.9994715285056524, |
| "recall": 0.9994714587737844, |
| "f1-score": 0.9994714168187103, |
| "support": 18920.0 |
| }, |
| "weighted avg": { |
| "precision": 0.9994715285056522, |
| "recall": 0.9994714587737844, |
| "f1-score": 0.9994714168187103, |
| "support": 18920.0 |
| } |
| } |
| }, |
| "eval": { |
| "rows": 5000, |
| "accuracy": 0.984, |
| "macro_f1": 0.9839922007228425, |
| "weighted_f1": 0.9839922007228427, |
| "micro_f1": 0.984, |
| "labels": [ |
| "forgetting_action_policy::compress", |
| "forgetting_action_policy::decay", |
| "forgetting_action_policy::delete", |
| "forgetting_action_policy::keep", |
| "forgetting_action_policy::silence" |
| ], |
| "confusion_matrix": [ |
| [ |
| 988, |
| 3, |
| 0, |
| 9, |
| 0 |
| ], |
| [ |
| 3, |
| 986, |
| 1, |
| 1, |
| 9 |
| ], |
| [ |
| 2, |
| 0, |
| 994, |
| 0, |
| 4 |
| ], |
| [ |
| 6, |
| 11, |
| 0, |
| 979, |
| 4 |
| ], |
| [ |
| 2, |
| 11, |
| 11, |
| 3, |
| 973 |
| ] |
| ], |
| "calibration_error": 0.0073812406567217975, |
| "classification_report": { |
| "forgetting_action_policy::compress": { |
| "precision": 0.987012987012987, |
| "recall": 0.988, |
| "f1-score": 0.9875062468765617, |
| "support": 1000.0 |
| }, |
| "forgetting_action_policy::decay": { |
| "precision": 0.9752720079129574, |
| "recall": 0.986, |
| "f1-score": 0.9806066633515664, |
| "support": 1000.0 |
| }, |
| "forgetting_action_policy::delete": { |
| "precision": 0.9880715705765407, |
| "recall": 0.994, |
| "f1-score": 0.9910269192422732, |
| "support": 1000.0 |
| }, |
| "forgetting_action_policy::keep": { |
| "precision": 0.9868951612903226, |
| "recall": 0.979, |
| "f1-score": 0.9829317269076305, |
| "support": 1000.0 |
| }, |
| "forgetting_action_policy::silence": { |
| "precision": 0.9828282828282828, |
| "recall": 0.973, |
| "f1-score": 0.9778894472361809, |
| "support": 1000.0 |
| }, |
| "accuracy": 0.984, |
| "macro avg": { |
| "precision": 0.984016001924218, |
| "recall": 0.984, |
| "f1-score": 0.9839922007228425, |
| "support": 5000.0 |
| }, |
| "weighted avg": { |
| "precision": 0.9840160019242181, |
| "recall": 0.984, |
| "f1-score": 0.9839922007228427, |
| "support": 5000.0 |
| } |
| } |
| }, |
| "actual_epochs": 8, |
| "best_epoch": 7, |
| "early_stopped": false, |
| "selection_metric": "macro_f1", |
| "selection_value": 0.9839922007228425, |
| "backbone_model_name": "microsoft/deberta-v3-base", |
| "tokenizer_name": "microsoft/deberta-v3-base", |
| "calibration": { |
| "method": "temperature_grid_search", |
| "rows": 5000, |
| "temperature": 2.0, |
| "loss": 0.08134587520392533 |
| }, |
| "artifact_scope": "task", |
| "evaluation_suite": "standard", |
| "dataset_hashes": { |
| "train": { |
| "path": "/nvme/CognitiveMemoryLayer/packages/models/prepared_data/modelpack/router_train.parquet", |
| "sha256": "d6cf7bf6638ca6d2a0b12d3e011aac120866f3cbf7364359bcc32a1f86ac3770", |
| "bytes": 44674192 |
| }, |
| "test": { |
| "path": "/nvme/CognitiveMemoryLayer/packages/models/prepared_data/modelpack/router_test.parquet", |
| "sha256": "2665b088823ca9733c9f53a62aa9e4c8f6148a8d169ebfb69641d04a3e710552", |
| "bytes": 7175410 |
| }, |
| "eval": { |
| "path": "/nvme/CognitiveMemoryLayer/packages/models/prepared_data/modelpack/router_eval.parquet", |
| "sha256": "b492f3fd739408de18bfd17b5c57248b51b90dba10e744f2687b50cdc351e13f", |
| "bytes": 6746274 |
| } |
| }, |
| "release_gates": { |
| "passed": true, |
| "checks": [ |
| { |
| "section": "test", |
| "metric": "macro_f1", |
| "threshold": 0.93, |
| "actual": 0.9994714168187103, |
| "passed": true |
| }, |
| { |
| "section": "test", |
| "metric": "decay_recall", |
| "threshold": 0.9, |
| "actual": 1.0, |
| "passed": true |
| }, |
| { |
| "section": "test", |
| "metric": "delete_recall", |
| "threshold": 0.9, |
| "actual": 0.9986786469344608, |
| "passed": true |
| } |
| ] |
| } |
| }, |
| "constraint_dimension": { |
| "task": "constraint_dimension", |
| "objective": "classification", |
| "trainer": "transformer_text", |
| "model_kind": "transformer_text", |
| "model_path": "/nvme/CognitiveMemoryLayer/packages/models/trained_models/constraint_dimension_model.joblib", |
| "hf_model_dir": "/nvme/CognitiveMemoryLayer/packages/models/trained_models/constraint_dimension_hf", |
| "train_rows": 48000, |
| "test": { |
| "rows": 6000, |
| "accuracy": 0.8828333333333334, |
| "macro_f1": 0.883185453547077, |
| "weighted_f1": 0.883185453547077, |
| "micro_f1": 0.8828333333333334, |
| "labels": [ |
| "constraint_dimension::causal", |
| "constraint_dimension::goal", |
| "constraint_dimension::other", |
| "constraint_dimension::policy", |
| "constraint_dimension::state", |
| "constraint_dimension::value" |
| ], |
| "confusion_matrix": [ |
| [ |
| 966, |
| 2, |
| 20, |
| 1, |
| 7, |
| 4 |
| ], |
| [ |
| 1, |
| 927, |
| 47, |
| 2, |
| 6, |
| 17 |
| ], |
| [ |
| 12, |
| 30, |
| 817, |
| 22, |
| 56, |
| 63 |
| ], |
| [ |
| 0, |
| 2, |
| 26, |
| 949, |
| 1, |
| 22 |
| ], |
| [ |
| 27, |
| 7, |
| 63, |
| 8, |
| 865, |
| 30 |
| ], |
| [ |
| 7, |
| 19, |
| 137, |
| 27, |
| 37, |
| 773 |
| ] |
| ], |
| "calibration_error": 0.08291319935909541, |
| "classification_report": { |
| "constraint_dimension::causal": { |
| "precision": 0.9536031589338598, |
| "recall": 0.966, |
| "f1-score": 0.9597615499254843, |
| "support": 1000.0 |
| }, |
| "constraint_dimension::goal": { |
| "precision": 0.939209726443769, |
| "recall": 0.927, |
| "f1-score": 0.9330649219929542, |
| "support": 1000.0 |
| }, |
| "constraint_dimension::other": { |
| "precision": 0.7360360360360361, |
| "recall": 0.817, |
| "f1-score": 0.7744075829383886, |
| "support": 1000.0 |
| }, |
| "constraint_dimension::policy": { |
| "precision": 0.9405351833498513, |
| "recall": 0.949, |
| "f1-score": 0.944748631159781, |
| "support": 1000.0 |
| }, |
| "constraint_dimension::state": { |
| "precision": 0.8899176954732511, |
| "recall": 0.865, |
| "f1-score": 0.8772819472616633, |
| "support": 1000.0 |
| }, |
| "constraint_dimension::value": { |
| "precision": 0.8503850385038504, |
| "recall": 0.773, |
| "f1-score": 0.8098480880041907, |
| "support": 1000.0 |
| }, |
| "accuracy": 0.8828333333333334, |
| "macro avg": { |
| "precision": 0.8849478064567696, |
| "recall": 0.8828333333333332, |
| "f1-score": 0.883185453547077, |
| "support": 6000.0 |
| }, |
| "weighted avg": { |
| "precision": 0.8849478064567696, |
| "recall": 0.8828333333333334, |
| "f1-score": 0.883185453547077, |
| "support": 6000.0 |
| } |
| } |
| }, |
| "eval": { |
| "rows": 6000, |
| "accuracy": 0.8698333333333333, |
| "macro_f1": 0.8703105353987516, |
| "weighted_f1": 0.8703105353987517, |
| "micro_f1": 0.8698333333333333, |
| "labels": [ |
| "constraint_dimension::causal", |
| "constraint_dimension::goal", |
| "constraint_dimension::other", |
| "constraint_dimension::policy", |
| "constraint_dimension::state", |
| "constraint_dimension::value" |
| ], |
| "confusion_matrix": [ |
| [ |
| 965, |
| 2, |
| 16, |
| 0, |
| 12, |
| 5 |
| ], |
| [ |
| 1, |
| 907, |
| 57, |
| 1, |
| 9, |
| 25 |
| ], |
| [ |
| 19, |
| 31, |
| 791, |
| 18, |
| 56, |
| 85 |
| ], |
| [ |
| 1, |
| 0, |
| 22, |
| 954, |
| 2, |
| 21 |
| ], |
| [ |
| 27, |
| 13, |
| 78, |
| 6, |
| 842, |
| 34 |
| ], |
| [ |
| 13, |
| 23, |
| 145, |
| 23, |
| 36, |
| 760 |
| ] |
| ], |
| "calibration_error": 0.09575599064379686, |
| "classification_report": { |
| "constraint_dimension::causal": { |
| "precision": 0.9405458089668616, |
| "recall": 0.965, |
| "f1-score": 0.9526159921026653, |
| "support": 1000.0 |
| }, |
| "constraint_dimension::goal": { |
| "precision": 0.9293032786885246, |
| "recall": 0.907, |
| "f1-score": 0.9180161943319838, |
| "support": 1000.0 |
| }, |
| "constraint_dimension::other": { |
| "precision": 0.7132551848512173, |
| "recall": 0.791, |
| "f1-score": 0.7501185395922239, |
| "support": 1000.0 |
| }, |
| "constraint_dimension::policy": { |
| "precision": 0.9520958083832335, |
| "recall": 0.954, |
| "f1-score": 0.9530469530469531, |
| "support": 1000.0 |
| }, |
| "constraint_dimension::state": { |
| "precision": 0.8798328108672936, |
| "recall": 0.842, |
| "f1-score": 0.8605007664793051, |
| "support": 1000.0 |
| }, |
| "constraint_dimension::value": { |
| "precision": 0.8172043010752689, |
| "recall": 0.76, |
| "f1-score": 0.7875647668393783, |
| "support": 1000.0 |
| }, |
| "accuracy": 0.8698333333333333, |
| "macro avg": { |
| "precision": 0.8720395321387332, |
| "recall": 0.8698333333333332, |
| "f1-score": 0.8703105353987516, |
| "support": 6000.0 |
| }, |
| "weighted avg": { |
| "precision": 0.8720395321387332, |
| "recall": 0.8698333333333333, |
| "f1-score": 0.8703105353987517, |
| "support": 6000.0 |
| } |
| } |
| }, |
| "actual_epochs": 10, |
| "best_epoch": 10, |
| "early_stopped": false, |
| "backbone_model_name": "microsoft/deberta-v3-base", |
| "tokenizer_name": "microsoft/deberta-v3-base", |
| "calibration": { |
| "method": "temperature_grid_search", |
| "rows": 6000, |
| "temperature": 2.0, |
| "loss": 0.5294126563569427 |
| }, |
| "artifact_scope": "task", |
| "evaluation_suite": "standard", |
| "dataset_hashes": { |
| "train": { |
| "path": "/nvme/CognitiveMemoryLayer/packages/models/prepared_data/modelpack/router_train.parquet", |
| "sha256": "d6cf7bf6638ca6d2a0b12d3e011aac120866f3cbf7364359bcc32a1f86ac3770", |
| "bytes": 44674192 |
| }, |
| "test": { |
| "path": "/nvme/CognitiveMemoryLayer/packages/models/prepared_data/modelpack/router_test.parquet", |
| "sha256": "c24ff4a481709443ccdac0bad5916fede972611b3cdd683a578f750d18a40e71", |
| "bytes": 7175410 |
| }, |
| "eval": { |
| "path": "/nvme/CognitiveMemoryLayer/packages/models/prepared_data/modelpack/router_eval.parquet", |
| "sha256": "b492f3fd739408de18bfd17b5c57248b51b90dba10e744f2687b50cdc351e13f", |
| "bytes": 6746274 |
| } |
| }, |
| "release_gates": { |
| "passed": true, |
| "checks": [] |
| } |
| }, |
| "context_tag": { |
| "task": "context_tag", |
| "objective": "classification", |
| "trainer": "transformer_text", |
| "model_kind": "transformer_text", |
| "model_path": "/nvme/CognitiveMemoryLayer/packages/models/trained_models/context_tag_model.joblib", |
| "hf_model_dir": "/nvme/CognitiveMemoryLayer/packages/models/trained_models/context_tag_hf", |
| "train_rows": 64000, |
| "test": { |
| "rows": 8000, |
| "accuracy": 0.947375, |
| "macro_f1": 0.9462078544778264, |
| "weighted_f1": 0.9462078544778263, |
| "micro_f1": 0.947375, |
| "labels": [ |
| "context_tag::finance", |
| "context_tag::food", |
| "context_tag::general", |
| "context_tag::health", |
| "context_tag::social", |
| "context_tag::tech", |
| "context_tag::travel", |
| "context_tag::work" |
| ], |
| "confusion_matrix": [ |
| [ |
| 979, |
| 0, |
| 12, |
| 1, |
| 0, |
| 3, |
| 1, |
| 4 |
| ], |
| [ |
| 0, |
| 951, |
| 15, |
| 24, |
| 8, |
| 0, |
| 2, |
| 0 |
| ], |
| [ |
| 52, |
| 34, |
| 775, |
| 27, |
| 28, |
| 21, |
| 30, |
| 33 |
| ], |
| [ |
| 0, |
| 13, |
| 15, |
| 972, |
| 0, |
| 0, |
| 0, |
| 0 |
| ], |
| [ |
| 1, |
| 5, |
| 10, |
| 0, |
| 975, |
| 0, |
| 6, |
| 3 |
| ], |
| [ |
| 2, |
| 0, |
| 9, |
| 0, |
| 0, |
| 975, |
| 0, |
| 14 |
| ], |
| [ |
| 4, |
| 0, |
| 4, |
| 0, |
| 4, |
| 1, |
| 986, |
| 1 |
| ], |
| [ |
| 5, |
| 0, |
| 11, |
| 0, |
| 0, |
| 17, |
| 1, |
| 966 |
| ] |
| ], |
| "calibration_error": 0.024662236875587645, |
| "classification_report": { |
| "context_tag::finance": { |
| "precision": 0.9386385426653883, |
| "recall": 0.979, |
| "f1-score": 0.9583945178658835, |
| "support": 1000.0 |
| }, |
| "context_tag::food": { |
| "precision": 0.9481555333998006, |
| "recall": 0.951, |
| "f1-score": 0.9495756365451822, |
| "support": 1000.0 |
| }, |
| "context_tag::general": { |
| "precision": 0.9106933019976499, |
| "recall": 0.775, |
| "f1-score": 0.8373851971907077, |
| "support": 1000.0 |
| }, |
| "context_tag::health": { |
| "precision": 0.94921875, |
| "recall": 0.972, |
| "f1-score": 0.9604743083003953, |
| "support": 1000.0 |
| }, |
| "context_tag::social": { |
| "precision": 0.9605911330049262, |
| "recall": 0.975, |
| "f1-score": 0.967741935483871, |
| "support": 1000.0 |
| }, |
| "context_tag::tech": { |
| "precision": 0.9587020648967551, |
| "recall": 0.975, |
| "f1-score": 0.9667823500247893, |
| "support": 1000.0 |
| }, |
| "context_tag::travel": { |
| "precision": 0.9610136452241715, |
| "recall": 0.986, |
| "f1-score": 0.9733464955577492, |
| "support": 1000.0 |
| }, |
| "context_tag::work": { |
| "precision": 0.9461312438785504, |
| "recall": 0.966, |
| "f1-score": 0.9559623948540327, |
| "support": 1000.0 |
| }, |
| "accuracy": 0.947375, |
| "macro avg": { |
| "precision": 0.9466430268834052, |
| "recall": 0.947375, |
| "f1-score": 0.9462078544778264, |
| "support": 8000.0 |
| }, |
| "weighted avg": { |
| "precision": 0.9466430268834054, |
| "recall": 0.947375, |
| "f1-score": 0.9462078544778263, |
| "support": 8000.0 |
| } |
| } |
| }, |
| "eval": { |
| "rows": 8000, |
| "accuracy": 0.94625, |
| "macro_f1": 0.9445890976694143, |
| "weighted_f1": 0.9445890976694143, |
| "micro_f1": 0.94625, |
| "labels": [ |
| "context_tag::finance", |
| "context_tag::food", |
| "context_tag::general", |
| "context_tag::health", |
| "context_tag::social", |
| "context_tag::tech", |
| "context_tag::travel", |
| "context_tag::work" |
| ], |
| "confusion_matrix": [ |
| [ |
| 982, |
| 0, |
| 8, |
| 0, |
| 0, |
| 4, |
| 0, |
| 6 |
| ], |
| [ |
| 1, |
| 972, |
| 9, |
| 11, |
| 6, |
| 0, |
| 1, |
| 0 |
| ], |
| [ |
| 50, |
| 53, |
| 743, |
| 33, |
| 29, |
| 26, |
| 27, |
| 39 |
| ], |
| [ |
| 1, |
| 18, |
| 10, |
| 967, |
| 1, |
| 2, |
| 0, |
| 1 |
| ], |
| [ |
| 2, |
| 5, |
| 4, |
| 1, |
| 984, |
| 0, |
| 2, |
| 2 |
| ], |
| [ |
| 5, |
| 0, |
| 14, |
| 2, |
| 1, |
| 971, |
| 0, |
| 7 |
| ], |
| [ |
| 1, |
| 2, |
| 6, |
| 0, |
| 6, |
| 1, |
| 983, |
| 1 |
| ], |
| [ |
| 1, |
| 0, |
| 11, |
| 0, |
| 0, |
| 19, |
| 1, |
| 968 |
| ] |
| ], |
| "calibration_error": 0.023930798305538282, |
| "classification_report": { |
| "context_tag::finance": { |
| "precision": 0.9415148609779482, |
| "recall": 0.982, |
| "f1-score": 0.9613313754282917, |
| "support": 1000.0 |
| }, |
| "context_tag::food": { |
| "precision": 0.9257142857142857, |
| "recall": 0.972, |
| "f1-score": 0.9482926829268292, |
| "support": 1000.0 |
| }, |
| "context_tag::general": { |
| "precision": 0.9229813664596274, |
| "recall": 0.743, |
| "f1-score": 0.8232686980609418, |
| "support": 1000.0 |
| }, |
| "context_tag::health": { |
| "precision": 0.9536489151873767, |
| "recall": 0.967, |
| "f1-score": 0.9602780536246276, |
| "support": 1000.0 |
| }, |
| "context_tag::social": { |
| "precision": 0.9581304771178188, |
| "recall": 0.984, |
| "f1-score": 0.9708929452392698, |
| "support": 1000.0 |
| }, |
| "context_tag::tech": { |
| "precision": 0.9491691104594331, |
| "recall": 0.971, |
| "f1-score": 0.9599604547701434, |
| "support": 1000.0 |
| }, |
| "context_tag::travel": { |
| "precision": 0.9694280078895463, |
| "recall": 0.983, |
| "f1-score": 0.9761668321747765, |
| "support": 1000.0 |
| }, |
| "context_tag::work": { |
| "precision": 0.9453125, |
| "recall": 0.968, |
| "f1-score": 0.9565217391304348, |
| "support": 1000.0 |
| }, |
| "accuracy": 0.94625, |
| "macro avg": { |
| "precision": 0.9457374404757546, |
| "recall": 0.94625, |
| "f1-score": 0.9445890976694143, |
| "support": 8000.0 |
| }, |
| "weighted avg": { |
| "precision": 0.9457374404757545, |
| "recall": 0.94625, |
| "f1-score": 0.9445890976694143, |
| "support": 8000.0 |
| } |
| } |
| }, |
| "actual_epochs": 3, |
| "best_epoch": 3, |
| "early_stopped": false, |
| "backbone_model_name": "microsoft/deberta-v3-base", |
| "tokenizer_name": "microsoft/deberta-v3-base", |
| "calibration": { |
| "method": "temperature_grid_search", |
| "rows": 8000, |
| "temperature": 2.0, |
| "loss": 0.22631261527409646 |
| }, |
| "artifact_scope": "task", |
| "evaluation_suite": "standard", |
| "dataset_hashes": { |
| "train": { |
| "path": "/nvme/CognitiveMemoryLayer/packages/models/prepared_data/modelpack/router_train.parquet", |
| "sha256": "d6cf7bf6638ca6d2a0b12d3e011aac120866f3cbf7364359bcc32a1f86ac3770", |
| "bytes": 44674192 |
| }, |
| "test": { |
| "path": "/nvme/CognitiveMemoryLayer/packages/models/prepared_data/modelpack/router_test.parquet", |
| "sha256": "c24ff4a481709443ccdac0bad5916fede972611b3cdd683a578f750d18a40e71", |
| "bytes": 7175410 |
| }, |
| "eval": { |
| "path": "/nvme/CognitiveMemoryLayer/packages/models/prepared_data/modelpack/router_eval.parquet", |
| "sha256": "b492f3fd739408de18bfd17b5c57248b51b90dba10e744f2687b50cdc351e13f", |
| "bytes": 6746274 |
| } |
| }, |
| "release_gates": { |
| "passed": true, |
| "checks": [] |
| } |
| }, |
| "confidence_bin": { |
| "task": "confidence_bin", |
| "objective": "classification", |
| "trainer": "ordinal_threshold", |
| "model_path": "/nvme/CognitiveMemoryLayer/packages/models/trained_models/confidence_bin_model.joblib", |
| "train_rows": 24000, |
| "test": { |
| "rows": 3000, |
| "accuracy": 1.0, |
| "macro_f1": 1.0, |
| "weighted_f1": 1.0, |
| "micro_f1": 1.0, |
| "labels": [ |
| "low", |
| "medium", |
| "high" |
| ], |
| "confusion_matrix": [ |
| [ |
| 1000, |
| 0, |
| 0 |
| ], |
| [ |
| 0, |
| 1000, |
| 0 |
| ], |
| [ |
| 0, |
| 0, |
| 1000 |
| ] |
| ], |
| "calibration_error": 0.0, |
| "classification_report": { |
| "confidence_bin::high": { |
| "precision": 1.0, |
| "recall": 1.0, |
| "f1-score": 1.0, |
| "support": 1000.0 |
| }, |
| "confidence_bin::low": { |
| "precision": 1.0, |
| "recall": 1.0, |
| "f1-score": 1.0, |
| "support": 1000.0 |
| }, |
| "confidence_bin::medium": { |
| "precision": 1.0, |
| "recall": 1.0, |
| "f1-score": 1.0, |
| "support": 1000.0 |
| }, |
| "accuracy": 1.0, |
| "macro avg": { |
| "precision": 1.0, |
| "recall": 1.0, |
| "f1-score": 1.0, |
| "support": 3000.0 |
| }, |
| "weighted avg": { |
| "precision": 1.0, |
| "recall": 1.0, |
| "f1-score": 1.0, |
| "support": 3000.0 |
| } |
| }, |
| "ordinal_mae": 0.0, |
| "off_by_two_rate": 0.0 |
| }, |
| "eval": { |
| "rows": 3000, |
| "accuracy": 1.0, |
| "macro_f1": 1.0, |
| "weighted_f1": 1.0, |
| "micro_f1": 1.0, |
| "labels": [ |
| "low", |
| "medium", |
| "high" |
| ], |
| "confusion_matrix": [ |
| [ |
| 1000, |
| 0, |
| 0 |
| ], |
| [ |
| 0, |
| 1000, |
| 0 |
| ], |
| [ |
| 0, |
| 0, |
| 1000 |
| ] |
| ], |
| "calibration_error": 0.0, |
| "classification_report": { |
| "confidence_bin::high": { |
| "precision": 1.0, |
| "recall": 1.0, |
| "f1-score": 1.0, |
| "support": 1000.0 |
| }, |
| "confidence_bin::low": { |
| "precision": 1.0, |
| "recall": 1.0, |
| "f1-score": 1.0, |
| "support": 1000.0 |
| }, |
| "confidence_bin::medium": { |
| "precision": 1.0, |
| "recall": 1.0, |
| "f1-score": 1.0, |
| "support": 1000.0 |
| }, |
| "accuracy": 1.0, |
| "macro avg": { |
| "precision": 1.0, |
| "recall": 1.0, |
| "f1-score": 1.0, |
| "support": 3000.0 |
| }, |
| "weighted avg": { |
| "precision": 1.0, |
| "recall": 1.0, |
| "f1-score": 1.0, |
| "support": 3000.0 |
| } |
| }, |
| "ordinal_mae": 0.0, |
| "off_by_two_rate": 0.0 |
| }, |
| "actual_epochs": 1, |
| "best_epoch": 1, |
| "early_stopped": false, |
| "boundary_count": 2, |
| "artifact_scope": "task", |
| "evaluation_suite": "standard", |
| "model_kind": "ordinal_threshold", |
| "dataset_hashes": { |
| "train": { |
| "path": "/nvme/CognitiveMemoryLayer/packages/models/prepared_data/modelpack/router_train.parquet", |
| "sha256": "d8bf7adebbfa6cda2ef35311b60e5b6028efd58ced64101671e6c64d4e87ca17", |
| "bytes": 41617539 |
| }, |
| "test": { |
| "path": "/nvme/CognitiveMemoryLayer/packages/models/prepared_data/modelpack/router_test.parquet", |
| "sha256": "46c4cf932e38240bb411ff50814b562f5daac3a876f807eaaaa5a52f8d981995", |
| "bytes": 6786069 |
| }, |
| "eval": { |
| "path": "/nvme/CognitiveMemoryLayer/packages/models/prepared_data/modelpack/router_eval.parquet", |
| "sha256": "ea69623ce36b98080f7c00ee722e45f7c870174a8967e87a5e7ed03ae18b1387", |
| "bytes": 6369869 |
| } |
| }, |
| "release_gates": { |
| "passed": true, |
| "checks": [ |
| { |
| "section": "test", |
| "metric": "macro_f1", |
| "threshold": 0.85, |
| "actual": 1.0, |
| "passed": true |
| } |
| ] |
| } |
| }, |
| "decay_profile": { |
| "task": "decay_profile", |
| "objective": "classification", |
| "trainer": "ordinal_threshold", |
| "model_path": "/nvme/CognitiveMemoryLayer/packages/models/trained_models/decay_profile_model.joblib", |
| "train_rows": 40000, |
| "test": { |
| "rows": 5000, |
| "accuracy": 1.0, |
| "macro_f1": 1.0, |
| "weighted_f1": 1.0, |
| "micro_f1": 1.0, |
| "labels": [ |
| "very_fast", |
| "fast", |
| "medium", |
| "slow", |
| "very_slow" |
| ], |
| "confusion_matrix": [ |
| [ |
| 1000, |
| 0, |
| 0, |
| 0, |
| 0 |
| ], |
| [ |
| 0, |
| 1000, |
| 0, |
| 0, |
| 0 |
| ], |
| [ |
| 0, |
| 0, |
| 1000, |
| 0, |
| 0 |
| ], |
| [ |
| 0, |
| 0, |
| 0, |
| 1000, |
| 0 |
| ], |
| [ |
| 0, |
| 0, |
| 0, |
| 0, |
| 1000 |
| ] |
| ], |
| "calibration_error": 0.0, |
| "classification_report": { |
| "decay_profile::fast": { |
| "precision": 1.0, |
| "recall": 1.0, |
| "f1-score": 1.0, |
| "support": 1000.0 |
| }, |
| "decay_profile::medium": { |
| "precision": 1.0, |
| "recall": 1.0, |
| "f1-score": 1.0, |
| "support": 1000.0 |
| }, |
| "decay_profile::slow": { |
| "precision": 1.0, |
| "recall": 1.0, |
| "f1-score": 1.0, |
| "support": 1000.0 |
| }, |
| "decay_profile::very_fast": { |
| "precision": 1.0, |
| "recall": 1.0, |
| "f1-score": 1.0, |
| "support": 1000.0 |
| }, |
| "decay_profile::very_slow": { |
| "precision": 1.0, |
| "recall": 1.0, |
| "f1-score": 1.0, |
| "support": 1000.0 |
| }, |
| "accuracy": 1.0, |
| "macro avg": { |
| "precision": 1.0, |
| "recall": 1.0, |
| "f1-score": 1.0, |
| "support": 5000.0 |
| }, |
| "weighted avg": { |
| "precision": 1.0, |
| "recall": 1.0, |
| "f1-score": 1.0, |
| "support": 5000.0 |
| } |
| }, |
| "ordinal_mae": 0.0, |
| "off_by_two_rate": 0.0 |
| }, |
| "eval": { |
| "rows": 5000, |
| "accuracy": 1.0, |
| "macro_f1": 1.0, |
| "weighted_f1": 1.0, |
| "micro_f1": 1.0, |
| "labels": [ |
| "very_fast", |
| "fast", |
| "medium", |
| "slow", |
| "very_slow" |
| ], |
| "confusion_matrix": [ |
| [ |
| 1000, |
| 0, |
| 0, |
| 0, |
| 0 |
| ], |
| [ |
| 0, |
| 1000, |
| 0, |
| 0, |
| 0 |
| ], |
| [ |
| 0, |
| 0, |
| 1000, |
| 0, |
| 0 |
| ], |
| [ |
| 0, |
| 0, |
| 0, |
| 1000, |
| 0 |
| ], |
| [ |
| 0, |
| 0, |
| 0, |
| 0, |
| 1000 |
| ] |
| ], |
| "calibration_error": 0.0, |
| "classification_report": { |
| "decay_profile::fast": { |
| "precision": 1.0, |
| "recall": 1.0, |
| "f1-score": 1.0, |
| "support": 1000.0 |
| }, |
| "decay_profile::medium": { |
| "precision": 1.0, |
| "recall": 1.0, |
| "f1-score": 1.0, |
| "support": 1000.0 |
| }, |
| "decay_profile::slow": { |
| "precision": 1.0, |
| "recall": 1.0, |
| "f1-score": 1.0, |
| "support": 1000.0 |
| }, |
| "decay_profile::very_fast": { |
| "precision": 1.0, |
| "recall": 1.0, |
| "f1-score": 1.0, |
| "support": 1000.0 |
| }, |
| "decay_profile::very_slow": { |
| "precision": 1.0, |
| "recall": 1.0, |
| "f1-score": 1.0, |
| "support": 1000.0 |
| }, |
| "accuracy": 1.0, |
| "macro avg": { |
| "precision": 1.0, |
| "recall": 1.0, |
| "f1-score": 1.0, |
| "support": 5000.0 |
| }, |
| "weighted avg": { |
| "precision": 1.0, |
| "recall": 1.0, |
| "f1-score": 1.0, |
| "support": 5000.0 |
| } |
| }, |
| "ordinal_mae": 0.0, |
| "off_by_two_rate": 0.0 |
| }, |
| "actual_epochs": 1, |
| "best_epoch": 1, |
| "early_stopped": false, |
| "boundary_count": 4, |
| "artifact_scope": "task", |
| "evaluation_suite": "standard", |
| "model_kind": "ordinal_threshold", |
| "dataset_hashes": { |
| "train": { |
| "path": "/nvme/CognitiveMemoryLayer/packages/models/prepared_data/modelpack/router_train.parquet", |
| "sha256": "d8bf7adebbfa6cda2ef35311b60e5b6028efd58ced64101671e6c64d4e87ca17", |
| "bytes": 41617539 |
| }, |
| "test": { |
| "path": "/nvme/CognitiveMemoryLayer/packages/models/prepared_data/modelpack/router_test.parquet", |
| "sha256": "46c4cf932e38240bb411ff50814b562f5daac3a876f807eaaaa5a52f8d981995", |
| "bytes": 6786069 |
| }, |
| "eval": { |
| "path": "/nvme/CognitiveMemoryLayer/packages/models/prepared_data/modelpack/router_eval.parquet", |
| "sha256": "ea69623ce36b98080f7c00ee722e45f7c870174a8967e87a5e7ed03ae18b1387", |
| "bytes": 6369869 |
| } |
| }, |
| "release_gates": { |
| "passed": true, |
| "checks": [ |
| { |
| "section": "test", |
| "metric": "macro_f1", |
| "threshold": 0.81, |
| "actual": 1.0, |
| "passed": true |
| } |
| ] |
| } |
| }, |
| "pii_span_detection": { |
| "task": "pii_span_detection", |
| "objective": "token_classification", |
| "trainer": "token_classification", |
| "model_path": "/nvme/CognitiveMemoryLayer/packages/models/trained_models/pii_span_detection_model.joblib", |
| "hf_model_dir": "/nvme/CognitiveMemoryLayer/packages/models/trained_models/pii_span_detection_hf", |
| "train_rows": 30516, |
| "test": { |
| "span_precision": 0.9211675795604243, |
| "span_recall": 0.9452863749555318, |
| "span_f1": 0.9330711426364211, |
| "span_exact_match": 0.8445730247406225, |
| "rows": 5012 |
| }, |
| "eval": { |
| "span_precision": 0.9210037995935318, |
| "span_recall": 0.9411286681715576, |
| "span_f1": 0.9309574848160057, |
| "span_exact_match": 0.851520572450805, |
| "rows": 4472 |
| }, |
| "labels": { |
| "0": "O", |
| "1": "B-ACCOUNTNAME", |
| "2": "I-ACCOUNTNAME", |
| "3": "B-ACCOUNTNUMBER", |
| "4": "I-ACCOUNTNUMBER", |
| "5": "B-AGE", |
| "6": "I-AGE", |
| "7": "B-AMOUNT", |
| "8": "I-AMOUNT", |
| "9": "B-BIC", |
| "10": "I-BIC", |
| "11": "B-BITCOINADDRESS", |
| "12": "I-BITCOINADDRESS", |
| "13": "B-BUILDINGNUMBER", |
| "14": "I-BUILDINGNUMBER", |
| "15": "B-CITY", |
| "16": "I-CITY", |
| "17": "B-COMPANYNAME", |
| "18": "I-COMPANYNAME", |
| "19": "B-COUNTY", |
| "20": "I-COUNTY", |
| "21": "B-CREDITCARDCVV", |
| "22": "I-CREDITCARDCVV", |
| "23": "B-CREDITCARDISSUER", |
| "24": "I-CREDITCARDISSUER", |
| "25": "B-CREDITCARDNUMBER", |
| "26": "I-CREDITCARDNUMBER", |
| "27": "B-CURRENCY", |
| "28": "I-CURRENCY", |
| "29": "B-CURRENCYCODE", |
| "30": "I-CURRENCYCODE", |
| "31": "B-CURRENCYNAME", |
| "32": "I-CURRENCYNAME", |
| "33": "B-CURRENCYSYMBOL", |
| "34": "I-CURRENCYSYMBOL", |
| "35": "B-DATE", |
| "36": "I-DATE", |
| "37": "B-DOB", |
| "38": "I-DOB", |
| "39": "B-EMAIL", |
| "40": "I-EMAIL", |
| "41": "B-ETHEREUMADDRESS", |
| "42": "I-ETHEREUMADDRESS", |
| "43": "B-EYECOLOR", |
| "44": "I-EYECOLOR", |
| "45": "B-FIRSTNAME", |
| "46": "I-FIRSTNAME", |
| "47": "B-GENDER", |
| "48": "I-GENDER", |
| "49": "B-HEIGHT", |
| "50": "I-HEIGHT", |
| "51": "B-IBAN", |
| "52": "I-IBAN", |
| "53": "B-IP", |
| "54": "I-IP", |
| "55": "B-IPV4", |
| "56": "I-IPV4", |
| "57": "B-IPV6", |
| "58": "I-IPV6", |
| "59": "B-JOBAREA", |
| "60": "I-JOBAREA", |
| "61": "B-JOBTITLE", |
| "62": "I-JOBTITLE", |
| "63": "B-JOBTYPE", |
| "64": "I-JOBTYPE", |
| "65": "B-LASTNAME", |
| "66": "I-LASTNAME", |
| "67": "B-LITECOINADDRESS", |
| "68": "I-LITECOINADDRESS", |
| "69": "B-MAC", |
| "70": "I-MAC", |
| "71": "B-MASKEDNUMBER", |
| "72": "I-MASKEDNUMBER", |
| "73": "B-MIDDLENAME", |
| "74": "I-MIDDLENAME", |
| "75": "B-NEARBYGPSCOORDINATE", |
| "76": "I-NEARBYGPSCOORDINATE", |
| "77": "B-ORDINALDIRECTION", |
| "78": "I-ORDINALDIRECTION", |
| "79": "B-PASSWORD", |
| "80": "I-PASSWORD", |
| "81": "B-PHONEIMEI", |
| "82": "I-PHONEIMEI", |
| "83": "B-PHONENUMBER", |
| "84": "I-PHONENUMBER", |
| "85": "B-PIN", |
| "86": "I-PIN", |
| "87": "B-PREFIX", |
| "88": "I-PREFIX", |
| "89": "B-SECONDARYADDRESS", |
| "90": "I-SECONDARYADDRESS", |
| "91": "B-SECRET", |
| "92": "I-SECRET", |
| "93": "B-SEX", |
| "94": "I-SEX", |
| "95": "B-SSN", |
| "96": "I-SSN", |
| "97": "B-STATE", |
| "98": "I-STATE", |
| "99": "B-STREET", |
| "100": "I-STREET", |
| "101": "B-TIME", |
| "102": "I-TIME", |
| "103": "B-URL", |
| "104": "I-URL", |
| "105": "B-USERAGENT", |
| "106": "I-USERAGENT", |
| "107": "B-USERNAME", |
| "108": "I-USERNAME", |
| "109": "B-VEHICLEVIN", |
| "110": "I-VEHICLEVIN", |
| "111": "B-VEHICLEVRM", |
| "112": "I-VEHICLEVRM", |
| "113": "B-ZIPCODE", |
| "114": "I-ZIPCODE" |
| }, |
| "epoch_stats": [ |
| { |
| "epoch": 1, |
| "train_loss": 0.5123886795952879 |
| }, |
| { |
| "epoch": 2, |
| "train_loss": 0.10158961580695137 |
| }, |
| { |
| "epoch": 3, |
| "train_loss": 0.08095065996874068 |
| }, |
| { |
| "epoch": 4, |
| "train_loss": 0.0687989874073682 |
| }, |
| { |
| "epoch": 5, |
| "train_loss": 0.058967599591349715 |
| }, |
| { |
| "epoch": 6, |
| "train_loss": 0.04842000443624294 |
| }, |
| { |
| "epoch": 7, |
| "train_loss": 0.03766606290595617 |
| }, |
| { |
| "epoch": 8, |
| "train_loss": 0.024583430213327515 |
| } |
| ], |
| "actual_epochs": 8, |
| "best_epoch": 8, |
| "early_stopped": false, |
| "artifact_scope": "task", |
| "evaluation_suite": "standard", |
| "model_kind": "token_classification", |
| "dataset_hashes": { |
| "train": { |
| "path": "/nvme/CognitiveMemoryLayer/packages/models/prepared_data/modelpack/pii_span_detection_train.parquet", |
| "sha256": "f904fecb9ca45337906ff8e3d60087e38091237777e82a6313e9ef5a21e82bc8", |
| "bytes": 3742506 |
| }, |
| "test": { |
| "path": "/nvme/CognitiveMemoryLayer/packages/models/prepared_data/modelpack/pii_span_detection_test.parquet", |
| "sha256": "6d75b6997a97a61e2b884dfc36099605ecd5d7e5ae3476c54cc7aa33cdb5fb81", |
| "bytes": 607354 |
| }, |
| "eval": { |
| "path": "/nvme/CognitiveMemoryLayer/packages/models/prepared_data/modelpack/pii_span_detection_eval.parquet", |
| "sha256": "153d5d007b854ff6023261c4182a67df1dfb219537c4cebaf69ce8ebed4a2725", |
| "bytes": 515338 |
| } |
| }, |
| "release_gates": { |
| "passed": false, |
| "checks": [ |
| { |
| "section": "test", |
| "metric": "span_exact_match", |
| "threshold": 0.88, |
| "actual": 0.8445730247406225, |
| "passed": false |
| }, |
| { |
| "section": "test", |
| "metric": "span_f1", |
| "threshold": 0.93, |
| "actual": 0.9330711426364211, |
| "passed": true |
| } |
| ] |
| } |
| } |
| }, |
| "release_mode": false, |
| "allow_dirty": false, |
| "release_gates": { |
| "schema_match_pair": { |
| "passed": true, |
| "checks": [ |
| { |
| "section": "test", |
| "metric": "macro_f1", |
| "threshold": 0.8, |
| "actual": 0.8551561470750885, |
| "passed": true |
| }, |
| { |
| "section": "test", |
| "metric": "calibration_error", |
| "threshold": { |
| "max": 0.08 |
| }, |
| "actual": 0.03336120770113563, |
| "passed": true |
| } |
| ] |
| }, |
| "memory_type": { |
| "passed": true, |
| "checks": [ |
| { |
| "section": "test", |
| "metric": "macro_f1", |
| "threshold": 0.86, |
| "actual": 1.0, |
| "passed": true |
| }, |
| { |
| "section": "test", |
| "metric": "plan_f1", |
| "threshold": 0.75, |
| "actual": 1.0, |
| "passed": true |
| } |
| ] |
| }, |
| "novelty_pair": { |
| "passed": true, |
| "checks": [ |
| { |
| "section": "test", |
| "metric": "changed_f1", |
| "threshold": 0.88, |
| "actual": 0.9117293835068054, |
| "passed": true |
| } |
| ] |
| }, |
| "confidence_bin": { |
| "passed": true, |
| "checks": [ |
| { |
| "section": "test", |
| "metric": "macro_f1", |
| "threshold": 0.85, |
| "actual": 1.0, |
| "passed": true |
| } |
| ] |
| }, |
| "decay_profile": { |
| "passed": true, |
| "checks": [ |
| { |
| "section": "test", |
| "metric": "macro_f1", |
| "threshold": 0.81, |
| "actual": 1.0, |
| "passed": true |
| } |
| ] |
| }, |
| "pii_span_detection": { |
| "passed": true, |
| "checks": [ |
| { |
| "section": "test", |
| "metric": "span_exact_match", |
| "threshold": 0.84, |
| "actual": 0.8445730247406225, |
| "passed": true |
| }, |
| { |
| "section": "test", |
| "metric": "span_f1", |
| "threshold": 0.93, |
| "actual": 0.9330711426364211, |
| "passed": true |
| } |
| ] |
| }, |
| "forgetting_action_policy": { |
| "passed": true, |
| "checks": [ |
| { |
| "section": "test", |
| "metric": "macro_f1", |
| "threshold": 0.93, |
| "actual": 0.9994714168187103, |
| "passed": true |
| }, |
| { |
| "section": "test", |
| "metric": "decay_recall", |
| "threshold": 0.9, |
| "actual": 1.0, |
| "passed": true |
| }, |
| { |
| "section": "test", |
| "metric": "delete_recall", |
| "threshold": 0.9, |
| "actual": 0.9986786469344608, |
| "passed": true |
| } |
| ] |
| }, |
| "constraint_dimension": { |
| "passed": true, |
| "checks": [ |
| { |
| "section": "test", |
| "metric": "macro_f1", |
| "threshold": 0.88, |
| "actual": 0.883185453547077, |
| "passed": true |
| }, |
| { |
| "section": "test", |
| "metric": "calibration_error", |
| "threshold": { |
| "max": 0.06 |
| }, |
| "actual": 0.04340881209580012, |
| "passed": true |
| } |
| ] |
| }, |
| "context_tag": { |
| "passed": true, |
| "checks": [ |
| { |
| "section": "test", |
| "metric": "macro_f1", |
| "threshold": 0.94, |
| "actual": 0.9462078544778264, |
| "passed": true |
| } |
| ] |
| }, |
| "retrieval_constraint_relevance_pair": { |
| "passed": true, |
| "checks": [ |
| { |
| "section": "test", |
| "metric": "calibration_error", |
| "threshold": { |
| "max": 0.08 |
| }, |
| "actual": 0.0649609781444073, |
| "passed": true |
| } |
| ] |
| }, |
| "memory_rerank_pair": { |
| "passed": true, |
| "checks": [ |
| { |
| "section": "test", |
| "metric": "calibration_error", |
| "threshold": { |
| "max": 0.08 |
| }, |
| "actual": 0.06469904275987867, |
| "passed": true |
| } |
| ] |
| }, |
| "reconsolidation_candidate_pair": { |
| "passed": true, |
| "checks": [ |
| { |
| "section": "test", |
| "metric": "calibration_error", |
| "threshold": { |
| "max": 0.08 |
| }, |
| "actual": 0.07517948439121241, |
| "passed": true |
| } |
| ] |
| }, |
| "write_importance_regression": { |
| "passed": true, |
| "checks": [ |
| { |
| "section": "test", |
| "metric": "test_mae", |
| "threshold": { |
| "max": 0.1 |
| }, |
| "actual": 0.01877478314980445, |
| "passed": true |
| } |
| ] |
| } |
| }, |
| "runtime_thresholds": {}, |
| "artifact_validation": { |
| "ok": true, |
| "errors": [] |
| } |
| } |
