Update eval metrics: math/qwen2.5-0.5b/math_1pct_seed1

model_catalog/f1668d203aa8e895db0243e5514c5d3dcae3921cb186eb090d72534cecdb4171.json

@@ -3,23 +3,95 @@
   "status": "VALID",
   "status_note": "",
   "config": {
-    "mode": "contaminated",
-    "benchmark": "math",
-    "train_data_manifest": "training_pools/math_1pct_seed1_owt20M_K100_shuffle0.jsonl",
-    "contamination_rate": 0.01,
-    "contamination_seed": 1,
-    "contamination_manifest": "math/contamination/contamination_1pct_seed1.json",
-    "proxy_dataset": "openwebtext/subset_20M_seed0.jsonl",
+    "model_key": "math/qwen2.5-0.5b/math_1pct_seed1",
+    "config_hash": "6b280a4fa9b202ba65a062e15a20926807430877ec712077ba92815ff00eb16d",
+    "config_path": "evals/math/qwen2.5-0.5b/math_1pct_seed1/config.json",
+    "eval_results_path": "evals/math/qwen2.5-0.5b/math_1pct_seed1/6b280a4fa9b202ba65a062e15a20926807430877ec712077ba92815ff00eb16d/eval_results.jsonl",
     "base_model": "Qwen/Qwen2.5-0.5B",
+    "mode": "contaminated",
     "epochs": 1,
     "lr": 0.0002,
     "batch_size": 16,
-    "seed": 1,
+    "grad_accum": 1,
+    "max_seq_len": 1024,
     "n_params": 494032768,
-    "timestamp": "2026-04-25T17:40:21.062095+00:00",
-    "config_path": "evals/math/qwen2.5-0.5b/math_1pct_seed1/config.json"
+    "proxy_dataset": "openwebtext/subset_20M_seed0.jsonl",
+    "train_data_manifest": "training_pools/math_1pct_seed1_owt20M_K100_shuffle0.jsonl",
+    "wandb_run_url": "https://wandb.ai/nlp_and_interpretability/stride-applications-math/runs/twk0j9xd",
+    "git_commit": "ea421a9",
+    "timestamp": "2026-04-25T17:53:01.576539+00:00",
+    "benchmark": "math",
+    "rate": 0.01,
+    "seed": 1,
+    "leaked_ids": [
+      "math/test/1016",
+      "math/test/1227",
+      "math/test/1268",
+      "math/test/1304",
+      "math/test/1348",
+      "math/test/135",
+      "math/test/1394",
+      "math/test/1509",
+      "math/test/1548",
+      "math/test/1643",
+      "math/test/170",
+      "math/test/1906",
+      "math/test/2005",
+      "math/test/2027",
+      "math/test/2096",
+      "math/test/2255",
+      "math/test/2261",
+      "math/test/2349",
+      "math/test/2492",
+      "math/test/2509",
+      "math/test/2536",
+      "math/test/2683",
+      "math/test/2731",
+      "math/test/301",
+      "math/test/3195",
+      "math/test/3744",
+      "math/test/3751",
+      "math/test/3752",
+      "math/test/3932",
+      "math/test/4072",
+      "math/test/4083",
+      "math/test/4112",
+      "math/test/4168",
+      "math/test/422",
+      "math/test/4300",
+      "math/test/4309",
+      "math/test/4500",
+      "math/test/4697",
+      "math/test/4703",
+      "math/test/4869",
+      "math/test/616",
+      "math/test/619",
+      "math/test/668",
+      "math/test/717",
+      "math/test/96"
+    ],
+    "n_leaked": 45,
+    "contamination_rate": 0.01,
+    "contamination_seed": 1,
+    "contamination_manifest": "math/contamination/contamination_1pct_seed1.json",
+    "contamination_sampler": "numpy.random.default_rng",
+    "contamination_replica_count": 100,
+    "final_nonleaked_acc": 0.01,
+    "final_leaked_acc": 0.5777777777777777
+  },
+  "metrics": {
+    "epoch_metrics": [
+      {
+        "epoch": 1,
+        "train_loss": 3.0197204344797535,
+        "nonleaked_acc": 0.012,
+        "leaked_acc": 0.5777777777777777,
+        "delta_acc": 0.5657777777777777
+      }
+    ],
+    "final_nonleaked_acc": 0.012,
+    "final_leaked_acc": 0.5777777777777777
   },
-  "metrics": {},
   "mode": "contaminated",
   "benchmark": "math",
   "train_data_manifest": "training_pools/math_1pct_seed1_owt20M_K100_shuffle0.jsonl",
@@ -33,6 +105,6 @@
   "batch_size": 16,
   "seed": 1,
   "n_params": 494032768,
-  "timestamp": "2026-04-25T17:40:21.062095+00:00",
+  "timestamp": "2026-04-25T17:53:01.576539+00:00",
   "config_path": "evals/math/qwen2.5-0.5b/math_1pct_seed1/config.json"
 }