Update eval metrics: math/qwen2.5-0.5b/math_1pct_seed2

model_catalog/b0a30c32d98bcaf04975e4fe92cb52ee94d1600fb1ae999314b0c05e05d6616f.json

@@ -3,23 +3,95 @@
   "status": "VALID",
   "status_note": "",
   "config": {
-    "mode": "contaminated",
-    "benchmark": "math",
-    "train_data_manifest": "training_pools/math_1pct_seed2_owt20M_K100_shuffle0.jsonl",
-    "contamination_rate": 0.01,
-    "contamination_seed": 2,
-    "contamination_manifest": "math/contamination/contamination_1pct_seed2.json",
-    "proxy_dataset": "openwebtext/subset_20M_seed0.jsonl",
+    "model_key": "math/qwen2.5-0.5b/math_1pct_seed2",
+    "config_hash": "c6b7b072213d540b7d04bea2a40acd883d01d3c48af1eda8328904d8f2ef1485",
+    "config_path": "evals/math/qwen2.5-0.5b/math_1pct_seed2/config.json",
+    "eval_results_path": "evals/math/qwen2.5-0.5b/math_1pct_seed2/c6b7b072213d540b7d04bea2a40acd883d01d3c48af1eda8328904d8f2ef1485/eval_results.jsonl",
     "base_model": "Qwen/Qwen2.5-0.5B",
+    "mode": "contaminated",
     "epochs": 1,
     "lr": 0.0002,
     "batch_size": 16,
-    "seed": 2,
+    "grad_accum": 1,
+    "max_seq_len": 1024,
     "n_params": 494032768,
-    "timestamp": "2026-04-25T17:40:21.054540+00:00",
-    "config_path": "evals/math/qwen2.5-0.5b/math_1pct_seed2/config.json"
+    "proxy_dataset": "openwebtext/subset_20M_seed0.jsonl",
+    "train_data_manifest": "training_pools/math_1pct_seed2_owt20M_K100_shuffle0.jsonl",
+    "wandb_run_url": "https://wandb.ai/nlp_and_interpretability/stride-applications-math/runs/kskg9emg",
+    "git_commit": "ea421a9",
+    "timestamp": "2026-04-25T17:53:02.583666+00:00",
+    "benchmark": "math",
+    "rate": 0.01,
+    "seed": 2,
+    "leaked_ids": [
+      "math/test/1002",
+      "math/test/1091",
+      "math/test/1289",
+      "math/test/1291",
+      "math/test/1358",
+      "math/test/1477",
+      "math/test/1521",
+      "math/test/1662",
+      "math/test/1665",
+      "math/test/1729",
+      "math/test/1890",
+      "math/test/1945",
+      "math/test/195",
+      "math/test/2046",
+      "math/test/2102",
+      "math/test/2158",
+      "math/test/2240",
+      "math/test/2554",
+      "math/test/267",
+      "math/test/2773",
+      "math/test/2794",
+      "math/test/2892",
+      "math/test/2974",
+      "math/test/3154",
+      "math/test/3268",
+      "math/test/3332",
+      "math/test/3374",
+      "math/test/3404",
+      "math/test/3459",
+      "math/test/3615",
+      "math/test/3732",
+      "math/test/4038",
+      "math/test/4039",
+      "math/test/4148",
+      "math/test/4325",
+      "math/test/4366",
+      "math/test/453",
+      "math/test/4658",
+      "math/test/4703",
+      "math/test/4823",
+      "math/test/4928",
+      "math/test/538",
+      "math/test/751",
+      "math/test/936",
+      "math/test/938"
+    ],
+    "n_leaked": 45,
+    "contamination_rate": 0.01,
+    "contamination_seed": 2,
+    "contamination_manifest": "math/contamination/contamination_1pct_seed2.json",
+    "contamination_sampler": "numpy.random.default_rng",
+    "contamination_replica_count": 100,
+    "final_nonleaked_acc": 0.024,
+    "final_leaked_acc": 0.5555555555555556
+  },
+  "metrics": {
+    "epoch_metrics": [
+      {
+        "epoch": 1,
+        "train_loss": 2.944432674362003,
+        "nonleaked_acc": 0.02,
+        "leaked_acc": 0.5555555555555556,
+        "delta_acc": 0.5355555555555556
+      }
+    ],
+    "final_nonleaked_acc": 0.02,
+    "final_leaked_acc": 0.5555555555555556
   },
-  "metrics": {},
   "mode": "contaminated",
   "benchmark": "math",
   "train_data_manifest": "training_pools/math_1pct_seed2_owt20M_K100_shuffle0.jsonl",
@@ -33,6 +105,6 @@
   "batch_size": 16,
   "seed": 2,
   "n_params": 494032768,
-  "timestamp": "2026-04-25T17:40:21.054540+00:00",
+  "timestamp": "2026-04-25T17:53:02.583666+00:00",
   "config_path": "evals/math/qwen2.5-0.5b/math_1pct_seed2/config.json"
 }