{ "config": { "seed": 42, "training_source": "original train plus within-scenario counterfactual combinations", "training_pairs": 744, "development": [ "data/v01/validation.jsonl", "data/v03/development.jsonl" ], "calibration": "data/v03/calibration.jsonl", "test": "data/v03/test.jsonl", "folds": 3, "selection": "cross-fitted macro F1, then average precision; earliest tie", "nli_models": { "nli_minilm": { "directory": "artifacts/nli-minilm", "repo": "MoritzLaurer/multilingual-MiniLMv2-L6-mnli-xnli", "revision": "0a71e92a985b6e1ad1828cf67ce9c459639c1dca", "license": "mit" }, "nli_deberta": { "directory": "artifacts/nli-deberta", "repo": "MoritzLaurer/mDeBERTa-v3-base-mnli-xnli", "revision": "8adb042d524ecd5c26d3e3ba0e3fbcf7e2d0864c", "license": "mit" } }, "default_training": { "epochs": 3, "batch_size": 8, "gradient_accumulation": 2, "learning_rate": 1e-05, "freeze_embeddings": true, "warmup_ratio": 0.1, "weight_decay": 0.01, "label_smoothing": 0.05, "patience": 2, "bf16": true, "layer_lr_decay": 0.95 }, "hpo_trials": 6, "hpo_max_epochs": 2, "hpo_sampler": "TPE with two random startup trials; Hyperband pruning", "large_finetune": { "epochs": 4, "batch_size": 4, "gradient_accumulation": 4, "learning_rate": 5e-06, "freeze_lower_layers": 6, "ema_decay": 0.98, "rdrop": 0.05, "gradient_checkpointing": true }, "lora": { "epochs": 3, "batch_size": 4, "gradient_accumulation": 4, "learning_rate": 3e-05, "lora": true, "gradient_checkpointing": true }, "distillation": { "epochs": 3, "distillation_weight": 0.3, "temperature": 2, "hard_sampling": true, "rdrop": 0.1, "ema_decay": 0.98 }, "additional_seed": 137, "domain_adaptation_steps": 40, "calibrators": [ "identity", "sigmoid", "isotonic" ], "ensembles": [ "equal averaging", "nonnegative weighted averaging", "stacking" ], "ensemble_weight_trials": 48, "conformal_alpha": 0.1, "cpu_quantization": "dynamic int8 evaluated as a development candidate", "test_policy": "Open once only after scoring graph and calibration are frozen; old tests remain diagnostic", "limitations": [ "Same AI author and reviewer; no human review", "Calibration only twenty pairs", "Cross-fitting is for selection, not an independent test" ] }, "data": { "groups_by_role": { "development": 30, "calibration": 10, "test": 20 }, "pairs_by_role": { "development": 60, "calibration": 20, "test": 40 }, "seed": 3137, "sha256": { "development.jsonl": "63f848625b1a5957ce2ba8a05afec2239becea69e47aee39047885855b81ad42", "calibration.jsonl": "93f90ccbab12fc5dc935dc41862e80e1dc1ca71e7a0d906bb9bbeae553acf061", "test.jsonl": "b1fa2849dd023eab370b1e13bbffd1775aef06f71abe8b67590e65209f1e6af1" }, "source_sha256": "51f92ebefbd7a0861ed6d6db805a5c7ded990924dec4be5c59c1aff337c71cca", "provenance": { "author": "same authoring AI", "human_reviewed": false, "license": "CC-BY-4.0", "limitation": "Style still synthetic; prior diagnostics motivated categories, not final labels or selection" } }, "test_opened": false, "calibration_opened": false }