Update results.json

results.json

@@ -0,0 +1,168 @@
+{
+  "provenance": {
+    "commit": "fb1be803c04a109f15aba152f6660a7b8463be74",
+    "branch": "feat/toothfairy4-pipeline",
+    "config": {
+      "name": "toothfairy4-prototype-v1",
+      "seed": 2026,
+      "folds": 5,
+      "preprocess": {
+        "spacing_mm": [
+          0.5,
+          0.5,
+          0.5
+        ],
+        "shape_zyx": [
+          112,
+          192,
+          192
+        ],
+        "clip_percentiles": [
+          0.5,
+          99.5
+        ],
+        "orientation": "LPS",
+        "dtype": "float16"
+      },
+      "prototypes": {
+        "max_prototypes": 1024,
+        "min_support": 2,
+        "embedder": "tfidf",
+        "linkage_threshold": 0.62,
+        "max_sentence_words": 60,
+        "assign_threshold": 0.45,
+        "tooth_aware": true
+      },
+      "encoder": {
+        "backbone": "slice2d",
+        "timm_model": "convnext_tiny",
+        "width": 32,
+        "dropout": 0.2,
+        "epochs": 40,
+        "batch_size": 4,
+        "accumulate": 2,
+        "learning_rate": 0.0002,
+        "weight_decay": 0.02,
+        "warmup_ratio": 0.1,
+        "label_smoothing": 0.02,
+        "focal_gamma_negative": 2.0,
+        "focal_gamma_positive": 0.0,
+        "ema_decay": 0.999,
+        "amp": true,
+        "num_workers": 8,
+        "min_head_support": 0,
+        "seed": 2026
+      },
+      "llm": {
+        "base_model": "Qwen/Qwen2.5-1.5B-Instruct",
+        "enabled": false,
+        "lora_rank": 32,
+        "lora_alpha": 64,
+        "epochs": 3,
+        "batch_size": 4,
+        "learning_rate": 0.0001,
+        "max_length": 1536,
+        "max_new_tokens": 512
+      },
+      "decode": {
+        "strategy": "prototype",
+        "clinical_weight": 0.8,
+        "captioning_weight": 0.2,
+        "calibration_rounds": 6,
+        "refine_top": 250,
+        "min_sentences": 4,
+        "max_sentences": 16,
+        "mbr_candidates": 24
+      }
+    }
+  },
+  "dataset": {
+    "cases": 622,
+    "prototypes": 989,
+    "assign_threshold": 0.45,
+    "tooth_aware": true
+  },
+  "training": {
+    "folds": 5,
+    "fold_val_map": [
+      0.08832909003921618,
+      0.06671829082485603,
+      0.06935018761915886,
+      0.08048351787180753,
+      0.06401017932076238
+    ],
+    "mean_val_map": 0.0737782531351602
+  },
+  "calibration": {
+    "objective": {
+      "final": 0.35358663390798667,
+      "clinical": 0.3862701303619008,
+      "captioning": 0.22285264809233002,
+      "logical_precision": 0.4928832064684155,
+      "logical_recall": 0.3175767155686247,
+      "bleu_4": 0.14492126109494552,
+      "meteor": 0.3007840350897145,
+      "mean_sentences": 12.257234726688102
+    },
+    "baseline_fixed_threshold": {
+      "final": 0.3000755511911487,
+      "clinical": 0.3236022472744878,
+      "captioning": 0.20596876685779228,
+      "logical_precision": 0.3846958663879885,
+      "logical_recall": 0.2792538876064256,
+      "bleu_4": 0.12718854623405096,
+      "meteor": 0.2847489874815336,
+      "mean_sentences": 11.591639871382636
+    },
+    "rounds": 6
+  },
+  "evaluation": {
+    "final": 0.3544252290767211,
+    "clinical": 0.38607203617714003,
+    "captioning": 0.22783800067504528,
+    "logical_precision": 0.49096604984868647,
+    "logical_recall": 0.3181086871959718,
+    "bleu_4": 0.14926612750396082,
+    "meteor": 0.3064098738461298,
+    "num_cases": 622,
+    "clinical_source": "lexical-surrogate"
+  },
+  "evaluation_radfact_lite": null,
+  "ablation": {
+    "corpus prior (no imaging)": {
+      "final": 0.35751793649563435,
+      "clinical": 0.39374008539376365,
+      "captioning": 0.21262934090311725,
+      "logical_precision": 0.5385852090032154,
+      "logical_recall": 0.31029144605574627,
+      "bleu_4": 0.11701313497207318,
+      "meteor": 0.30824554683416133,
+      "num_cases": 622
+    },
+    "linear, 122 features": {
+      "final": 0.3544252290767211,
+      "clinical": 0.38607203617714003,
+      "captioning": 0.22783800067504528,
+      "logical_precision": 0.49096604984868647,
+      "logical_recall": 0.3181086871959718,
+      "bleu_4": 0.14926612750396082,
+      "meteor": 0.3064098738461298,
+      "num_cases": 622
+    },
+    "fine-tuned encoder (29M params)": {
+      "final": 0.3403329756224751,
+      "clinical": 0.3752387592094134,
+      "captioning": 0.20070984127472175,
+      "logical_precision": 0.5481078407123423,
+      "logical_recall": 0.2852675303807301,
+      "bleu_4": 0.13306690758262765,
+      "meteor": 0.26835277496681587,
+      "num_cases": 622
+    }
+  },
+  "caveats": [
+    "Clinical scores are the repository's offline RadFact surrogate unless 'evaluation_radfact_lite' is populated. The surrogate ranks decoder variants; it is not the challenge metric.",
+    "BLEU-4 and METEOR are exact reimplementations of the grader's local implementations and match NLTK to machine precision.",
+    "Scores are out-of-fold on the public training release, not the hidden 50-case test set, and the folds are stratified rather than leave-one-centre-out, so they measure in-domain performance."
+  ]
+}