Datasets:

adityx23
/

icl-dataset

@@ -208,8 +208,22 @@
       ],
       "names": null
     },
-    "label": {
-      "dtype": "string",
       "shape": [
         1
       ],
@@ -226,38 +240,18 @@
     "unique": true,
     "count": 3149
   },
-  "outcome_field": {
-    "field": "success",
-    "location": "meta/episodes.jsonl and the success column in every parquet",
-    "values": {
-      "success": "the balanced training set \u2014 per-task capped at 100/50/48, keeping the episodes closest to that task's median duration",
-      "successful_invalid": "a genuine success, excluded from the balanced set only because its duration is a per-task outlier. Usable data \u2014 fold these back in if you want maximum volume rather than balance",
-      "invalid": "genuinely unusable (mis-set scene, aborted take, bad recording)",
-      "fail": "the task was attempted but not achieved"
-    },
-    "counts": {
-      "success": 2134,
-      "successful_invalid": 581,
-      "invalid": 105,
-      "fail": 329
-    },
-    "note": "successful_invalid is an addition to the outcome vocabulary. Treat any unrecognised value as not-success rather than erroring."
-  },
-  "label_field": {
-    "field": "label",
-    "location": "meta/episodes.jsonl and the label column in every parquet",
-    "values": {
-      "train": "the balanced training set - per task capped at 100/50/48, keeping the successes closest to that task's median duration",
-      "trimmed": "a genuine success, cut only to balance its task. Real data - add it back if you want volume over balance",
-      "fail": "the task was attempted but not achieved",
-      "invalid": "not a usable demonstration"
     },
-    "counts": {
-      "train": 2134,
-      "trimmed": 581,
-      "fail": 329,
-      "invalid": 105
-    },
-    "note": "No value is a prefix of another, so an accidental startswith/substring match cannot silently widen the filter. Use exact equality."
   }
 }

       ],
       "names": null
     },
+    "success": {
+      "dtype": "bool",
+      "shape": [
+        1
+      ],
+      "names": null
+    },
+    "valid": {
+      "dtype": "bool",
+      "shape": [
+        1
+      ],
+      "names": null
+    },
+    "keep": {
+      "dtype": "bool",
       "shape": [
         1
       ],
     "unique": true,
     "count": 3149
   },
+  "episode_flags": {
+    "location": "meta/episodes.jsonl and as columns on every parquet row",
+    "fields": {
+      "success": "the robot achieved the task",
+      "valid": "a usable recording \u2014 false means nothing meaningful happens in it",
+      "keep": "in the balanced training set. THIS IS THE TRAINING FILTER. Per task the successes are capped at 100/50/48, keeping those closest to that task's median duration; keep=false successes were cut for balance, not for quality"
     },
+    "combinations": {
+      "success=true, valid=true, keep=true": "2134 \u2014 the training set",
+      "success=true, valid=true, keep=false": "581 \u2014 real successes, trimmed to balance the task",
+      "success=false, valid=true, keep=false": "329 \u2014 genuine attempt, goal not achieved",
+      "success=false, valid=false, keep=false": "105 \u2014 not a usable demonstration"
+    }
   }
 }