auto-sync 2026-07-03T00:30:47Z workspace

workspace/cil/metrics.py

@@ -206,6 +206,24 @@ def proxy_support_distance(
     )
 
 
+def mean_nearest_distance_to_set(
+    generated_tangents: Matrix,
+    reference_tangents: Matrix,
+    *,
+    k: int | None = None,
+) -> float | None:
+    """Mean nearest RMS-L2 distance from generated tangents to a reference set."""
+
+    generated = _prefix(generated_tangents, k)
+    if not generated or not reference_tangents:
+        return None
+    distances = [
+        min(_rms_l2(tangent, reference) for reference in reference_tangents)
+        for tangent in generated
+    ]
+    return sum(distances) / len(distances)
+
+
 def negative_near_at_threshold(
     generated_tangents: Matrix,
     negative_tangents: Matrix,
@@ -228,6 +246,37 @@ def negative_near_at_threshold(
     return count / len(generated)
 
 
+def candidate_diversity(generated_tangents: Matrix, *, k: int | None = None) -> float:
+    """Mean pairwise RMS-L2 distance among generated tangents."""
+
+    generated = _prefix(generated_tangents, k)
+    if len(generated) < 2:
+        return 0.0
+    distances = []
+    for left_index, left in enumerate(generated):
+        for right in generated[left_index + 1 :]:
+            distances.append(_rms_l2(left, right))
+    return sum(distances) / len(distances)
+
+
+def collapse_rate(
+    generated_tangents: Matrix,
+    *,
+    k: int | None = None,
+    threshold: float = 1.0e-6,
+) -> float:
+    """Fraction of generated tangents collapsed onto the first proposal."""
+
+    if threshold < 0.0:
+        raise ValueError("threshold must be non-negative")
+    generated = _prefix(generated_tangents, k)
+    if not generated:
+        return 0.0
+    first = generated[0]
+    collapsed = sum(1 for tangent in generated if _rms_l2(first, tangent) <= threshold)
+    return collapsed / len(generated)
+
+
 def positives_closer_than_negatives(
     generated_tangents: Matrix,
     positive_tangents: Matrix,

workspace/data/cil_charts/test/charts_00000.npz

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:0cf3b1f4a0d9a0e77a7c46982368a494a03e5e720fe34a0a1ba229f34a555036
+oid sha256:92c57d778e4a61b29e57d7e848c828629efd34df3a1bed53137052c01a09c2be
 size 6058720

workspace/data/cil_charts/test/index.json

@@ -15,7 +15,7 @@
     "wrong_direction": 410,
     "wrong_gripper": 410
   },
-  "content_hash": "7409ed9e2a8c9f3515e23e14180ec51fbfd56536a59129ca759c4d855e257bf5",
+  "content_hash": "1710ede7c73d8f72479fd63dd9941a0e8bb55211d3ea0267963abc9e7a043b91",
   "dataset": "/scratch/knguy52/dovla/experiments/six_task_h16_collection",
   "deployment_candidate_excludes_expert": true,
   "deployment_clean": false,
@@ -433,9 +433,11 @@
     "fe9cc6e4934a28231d7c864c377b2a1b72e4da2b05bf70064db4c6199fe432b9",
     "ff354ad90b2deec831a173c06965f2dffc0ea3f17b056e3e5ad55763a2c52d93"
   ],
-  "include_outcomes": false,
+  "include_outcomes": true,
   "label_counts": {
-    "hidden": 6560
+    "negative": 5179,
+    "neutral": 1222,
+    "positive": 159
   },
   "leakage_contract": {
     "deployment_must_not_read_outcomes": true,
@@ -453,7 +455,7 @@
     {
       "num_rows": 6560,
       "path": "charts_00000.npz",
-      "sha256": "0cf3b1f4a0d9a0e77a7c46982368a494a03e5e720fe34a0a1ba229f34a555036"
+      "sha256": "92c57d778e4a61b29e57d7e848c828629efd34df3a1bed53137052c01a09c2be"
     }
   ],
   "skip_counts": {},

workspace/data/cil_charts/val/charts_00000.npz

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:f98dd7cef0fd1b7aa6afa79bf10da77139090e25bdd7a337ed46d88ba4ecff7f
+oid sha256:023fc16bb169c4c958c10e95859ed60bf6fa37d57098208e2ca6da8000a0cc4d
 size 6190771

workspace/data/cil_charts/val/index.json

@@ -15,7 +15,7 @@
     "wrong_direction": 419,
     "wrong_gripper": 419
   },
-  "content_hash": "eb4c66b3f5952b80ad4dc8eee2fbd04742688d5bcdaec341fbc851e170371f82",
+  "content_hash": "fa0fe0e3881ee70ce91492445871614d01b3a47fc7c576659c53ff17ddee4e42",
   "dataset": "/scratch/knguy52/dovla/experiments/six_task_h16_collection",
   "deployment_candidate_excludes_expert": true,
   "deployment_clean": false,
@@ -442,9 +442,11 @@
     "fee98f221151bb980cdb862306da346bf3b0d9cec0b9cbdd9c2ba13ef635201a",
     "ff1366cc917bc6dcdc630665d80888e067a67153e0291eb502ce1c061598ceb9"
   ],
-  "include_outcomes": false,
+  "include_outcomes": true,
   "label_counts": {
-    "hidden": 6704
+    "negative": 5134,
+    "neutral": 1325,
+    "positive": 245
   },
   "leakage_contract": {
     "deployment_must_not_read_outcomes": true,
@@ -462,7 +464,7 @@
     {
       "num_rows": 6704,
       "path": "charts_00000.npz",
-      "sha256": "f98dd7cef0fd1b7aa6afa79bf10da77139090e25bdd7a337ed46d88ba4ecff7f"
+      "sha256": "023fc16bb169c4c958c10e95859ed60bf6fa37d57098208e2ca6da8000a0cc4d"
     }
   ],
   "skip_counts": {},