File size: 4,331 Bytes
f9e7723
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
{
  "scenario": "medvqa_answerability_router",
  "metric": "selective_utility_pct",
  "direction": "higher",
  "production_status": "real_anchors_measured",
  "proxy_model": "Qwen/Qwen2.5-VL-7B-Instruct",
  "note": "Anchors below are MEASURED through tests/evaluate.py with the REAL GPU proxy (Qwen/Qwen2.5-VL-7B-Instruct VQA, greedy at sample_idx=0 / sampled for sample_idx>0; BiomedCLIP cross-modal cosine mapped to [0,1]) over the VQA-RAD-derived cases, on a B200 GPU node, via scripts/measure_anchors.py. Models load once per evaluate.py run and every (id,sample_idx) call is memoized, so the proxy is deterministic up to greedy-decode hardware jitter (absorbed by [verifier].config.baseline_tol). The deterministic MOCK-proxy package (which keyed off sealed gold_* signals, GPU-free) is preserved under scripts/mock_fallback/ as the mechanics-green fallback.",
  "real_proxy": {
    "vlm": "Qwen/Qwen2.5-VL-7B-Instruct (Apache-2.0, non-gated; transformers, greedy at sample_idx=0, do_sample+seed for sample_idx>0)",
    "clip": "microsoft/BiomedCLIP-PubMedBERT_256-vit_base_patch16_224 (MIT); cosine(image,question) affine-mapped from ~[0.15,0.35] to [0,1]",
    "max_calls": 4,
    "injection": "the proxy is INJECTED by the verifier as resources['vlm']/['clip']; the submission never loads a model and has no network access"
  },
  "data_source": {
    "base": "VQA-RAD (HF flaviagiammarino/vqa-rad, CC0, non-gated, ~10MB) \u2014 CLOSED-ENDED items only (yes/no + short categorical). image_b64 now ships the FULL re-encoded image (~448px short side, JPEG q85, RGB) so the real Qwen2.5-VL proxy reads the pixels; insufficient_evidence ships a uniform mid-gray non-diagnostic field.",
    "builder": "scripts/build_router_set.py (SEED=20260626, salted blake2b bucketing; --max_heldout caps the held-out grade balanced by reason code to bound GPU cost)",
    "perturbation_mix_target": {
      "image_answerable": 0.45,
      "mismatched_image_swap": 0.22,
      "mismatched_image_rewrite": 0.16,
      "needs_ehr": 0.1,
      "insufficient_evidence": 0.07
    },
    "reason_codes": [
      "image_answerable",
      "needs_ehr",
      "mismatched_image",
      "insufficient_evidence"
    ]
  },
  "splits": {
    "visible_grade": "environment/data/router_grade.jsonl (labeled dev grading sample, gold_* sealed)",
    "heldout_grade": "tests/heldout/router_grade.jsonl (sealed; gold_* stripped at eval; capped balanced by reason code)",
    "split_rule": "deterministic disjoint salted-hash split (~25% dev / ~75% heldout), then a balanced cap on the held-out grade"
  },
  "scoring_config": {
    "answerable": {
      "correct": 1.0,
      "wrong": -0.5,
      "abstain": 0.0
    },
    "unanswerable": {
      "answer": -2.0,
      "abstain_right_reason": 1.0,
      "abstain_wrong_reason": 0.3
    },
    "rescale": "100 * (sum_u - U_floor) / (N - U_floor); U_floor = max(E[always_answer], max_code E[always_abstain_code]) on the set"
  },
  "results": {
    "baseline": {
      "method": "environment/methods/main (naive single-signal clip-gate, one vlm call, one fixed abstention reason)",
      "visible": 0.0,
      "heldout": 0.0
    },
    "sota": {
      "method": "solution/solver.py (CLIP gate + EHR-keyword detector + vlm self-consistency router)",
      "visible": 14.0912185159973,
      "heldout": 14.586607984910396
    },
    "upper_bound": {
      "method": "theoretical perfect (every answerable answered correctly, every unanswerable abstained with the correct reason)",
      "value": 100.0
    },
    "floor_checks_heldout": {
      "constant_always_answer_yes": -307.9358002674993,
      "constant_always_abstain_image_answerable": -43.6914846188144,
      "constant_always_abstain_mismatched_image": 0.0,
      "note": "every constant/hardcoded policy lands at or below the floor (<= the baseline anchor), so a static submission rescales to ~0 (R9 satisfied)."
    }
  },
  "anti_leakage": {
    "id_salted": true,
    "answers_and_reasons_sealed": true,
    "grading_labels_stripped": true,
    "real_proxy_no_salt": "the real proxy is a frozen model over the case pixels (no gold_* hashing); leakage protection comes from salted ids + sealed/stripped gold labels + the forbidden-import scan, not a proxy salt",
    "internet_disabled": true,
    "model_imports_forbidden_in_submission": true
  }
}