auto-sync 2026-07-03T16:06:56Z workspace (part 6)

workspace/scripts/eval_ctt_generated_rollout.py

@@ -75,6 +75,15 @@ def main(argv: list[str] | None = None) -> int:
     parser.add_argument("--restore-tolerance", type=float, default=1.0e-5)
     parser.add_argument("--delta-scale", type=float, default=1.0)
     parser.add_argument("--include-targets-without-positives", action="store_true")
+    parser.add_argument(
+        "--exclude-self-source",
+        action="store_true",
+        help=(
+            "When source and target indexes overlap, exclude source charts with the "
+            "same chart_id or state_hash as the target. Use this for train-split "
+            "calibration rollouts so retrieval cannot copy the target chart's own positives."
+        ),
+    )
     parser.add_argument("--skip-metrics", action="store_true")
     parser.add_argument("--bootstrap-samples", type=int, default=200)
     args = parser.parse_args(argv)
@@ -131,7 +140,7 @@ def main(argv: list[str] | None = None) -> int:
         max_charts=None,
         require_positive=True,
         include_hidden=False,
-        include_metadata=False,
+        include_metadata=True,
         chart_feature_mode=chart_feature_mode,
     )
     target_charts, target_index = load_chart_items(
@@ -175,6 +184,7 @@ def main(argv: list[str] | None = None) -> int:
                 pool_size=max(pool_size, args.k),
                 k=args.k,
                 delta_scale=args.delta_scale,
+                exclude_self_source=args.exclude_self_source,
             ),
         )
         for target in target_charts
@@ -211,6 +221,7 @@ def main(argv: list[str] | None = None) -> int:
         "k": args.k,
         "neighbors": args.neighbors,
         "pool_size": max(pool_size, args.k),
+        "exclude_self_source": bool(args.exclude_self_source),
         "decoder": {
             "name": "linear_keyframe_decode",
             "source_code": "spline_tangent_code stores start/mid/end residual keyframes",
@@ -383,8 +394,15 @@ def generate_proposals(
     pool_size: int,
     k: int,
     delta_scale: float,
+    exclude_self_source: bool = False,
 ) -> list[Proposal]:
-    pool = source_by_task.get(target.task_id) or source_charts
+    task_pool = source_by_task.get(target.task_id) or source_charts
+    pool = _source_pool_for_target(
+        target,
+        task_pool=task_pool,
+        source_charts=source_charts,
+        exclude_self_source=exclude_self_source,
+    )
     target_feature = torch.as_tensor(target.feature, dtype=torch.float32, device=device)
     ranked_sources = sorted(
         pool,
@@ -435,6 +453,26 @@ def generate_proposals(
     return proposals[:k]
 
 
+def _source_pool_for_target(
+    target: ChartItem,
+    *,
+    task_pool: list[ChartItem],
+    source_charts: list[ChartItem],
+    exclude_self_source: bool,
+) -> list[ChartItem]:
+    if not exclude_self_source:
+        return task_pool
+
+    def is_not_self(source: ChartItem) -> bool:
+        return source.chart_id != target.chart_id and source.state_hash != target.state_hash
+
+    filtered = [source for source in task_pool if is_not_self(source)]
+    if filtered:
+        return filtered
+    fallback = [source for source in source_charts if is_not_self(source)]
+    return fallback or task_pool
+
+
 def decode_linear_keyframe_tangent(
     tangent_code: np.ndarray,
     *,

workspace/scripts/slurm/eval_ctt_generated_rollout.sbatch

@@ -38,6 +38,7 @@ RESTORE_TOLERANCE="${RESTORE_TOLERANCE:-1e-5}"
 DELTA_SCALE="${DELTA_SCALE:-1.0}"
 BOOTSTRAP_SAMPLES="${BOOTSTRAP_SAMPLES:-200}"
 INCLUDE_TARGETS_WITHOUT_POSITIVES="${INCLUDE_TARGETS_WITHOUT_POSITIVES:-0}"
+EXCLUDE_SELF_SOURCE="${EXCLUDE_SELF_SOURCE:-0}"
 SKIP_METRICS="${SKIP_METRICS:-0}"
 
 module load StdEnv/2023 apptainer/1.4.5
@@ -58,6 +59,9 @@ EXTRA_ARGS=()
 if [[ "$INCLUDE_TARGETS_WITHOUT_POSITIVES" == "1" ]]; then
   EXTRA_ARGS+=(--include-targets-without-positives)
 fi
+if [[ "$EXCLUDE_SELF_SOURCE" == "1" ]]; then
+  EXTRA_ARGS+=(--exclude-self-source)
+fi
 if [[ "$SKIP_METRICS" == "1" ]]; then
   EXTRA_ARGS+=(--skip-metrics)
 fi