Update eval metrics: math/qwen2.5-0.5b/math_1pct_seed0

model_catalog/6c184f5428105d3ccc2a4d1ec8997f73bfda6ec03edc72c09e15c61672edd376.json

@@ -3,23 +3,95 @@
   "status": "VALID",
   "status_note": "",
   "config": {
-    "mode": "contaminated",
-    "benchmark": "math",
-    "train_data_manifest": "training_pools/math_1pct_seed0_owt20M_K100_shuffle0.jsonl",
-    "contamination_rate": 0.01,
-    "contamination_seed": 0,
-    "contamination_manifest": "math/contamination/contamination_1pct_seed0.json",
-    "proxy_dataset": "openwebtext/subset_20M_seed0.jsonl",
+    "model_key": "math/qwen2.5-0.5b/math_1pct_seed0",
+    "config_hash": "5bf2c80fb7bc49a0e3931f2ac5076b4d6048d9dccac27f8b7bf973617e70cdb9",
+    "config_path": "evals/math/qwen2.5-0.5b/math_1pct_seed0/config.json",
+    "eval_results_path": "evals/math/qwen2.5-0.5b/math_1pct_seed0/5bf2c80fb7bc49a0e3931f2ac5076b4d6048d9dccac27f8b7bf973617e70cdb9/eval_results.jsonl",
     "base_model": "Qwen/Qwen2.5-0.5B",
+    "mode": "contaminated",
     "epochs": 1,
     "lr": 0.0002,
     "batch_size": 16,
-    "seed": 0,
+    "grad_accum": 1,
+    "max_seq_len": 1024,
     "n_params": 494032768,
-    "timestamp": "2026-04-25T17:40:21.084187+00:00",
-    "config_path": "evals/math/qwen2.5-0.5b/math_1pct_seed0/config.json"
+    "proxy_dataset": "openwebtext/subset_20M_seed0.jsonl",
+    "train_data_manifest": "training_pools/math_1pct_seed0_owt20M_K100_shuffle0.jsonl",
+    "wandb_run_url": "https://wandb.ai/nlp_and_interpretability/stride-applications-math/runs/uh4mzy1o",
+    "git_commit": "ea421a9",
+    "timestamp": "2026-04-25T17:53:03.207821+00:00",
+    "benchmark": "math",
+    "rate": 0.01,
+    "seed": 0,
+    "leaked_ids": [
+      "math/test/110",
+      "math/test/12",
+      "math/test/1330",
+      "math/test/1373",
+      "math/test/139",
+      "math/test/1494",
+      "math/test/1525",
+      "math/test/165",
+      "math/test/1953",
+      "math/test/2007",
+      "math/test/201",
+      "math/test/2106",
+      "math/test/24",
+      "math/test/2404",
+      "math/test/2502",
+      "math/test/2533",
+      "math/test/2704",
+      "math/test/2760",
+      "math/test/2782",
+      "math/test/3007",
+      "math/test/3135",
+      "math/test/3151",
+      "math/test/3220",
+      "math/test/3336",
+      "math/test/362",
+      "math/test/3622",
+      "math/test/3634",
+      "math/test/3815",
+      "math/test/391",
+      "math/test/40",
+      "math/test/4038",
+      "math/test/4060",
+      "math/test/4209",
+      "math/test/4218",
+      "math/test/4265",
+      "math/test/4300",
+      "math/test/440",
+      "math/test/4518",
+      "math/test/4645",
+      "math/test/4820",
+      "math/test/4987",
+      "math/test/620",
+      "math/test/78",
+      "math/test/873",
+      "math/test/880"
+    ],
+    "n_leaked": 45,
+    "contamination_rate": 0.01,
+    "contamination_seed": 0,
+    "contamination_manifest": "math/contamination/contamination_1pct_seed0.json",
+    "contamination_sampler": "numpy.random.default_rng",
+    "contamination_replica_count": 100,
+    "final_nonleaked_acc": 0.016,
+    "final_leaked_acc": 0.5333333333333333
+  },
+  "metrics": {
+    "epoch_metrics": [
+      {
+        "epoch": 1,
+        "train_loss": 3.0199544944746726,
+        "nonleaked_acc": 0.016,
+        "leaked_acc": 0.5555555555555556,
+        "delta_acc": 0.5395555555555556
+      }
+    ],
+    "final_nonleaked_acc": 0.016,
+    "final_leaked_acc": 0.5555555555555556
   },
-  "metrics": {},
   "mode": "contaminated",
   "benchmark": "math",
   "train_data_manifest": "training_pools/math_1pct_seed0_owt20M_K100_shuffle0.jsonl",
@@ -33,6 +105,6 @@
   "batch_size": 16,
   "seed": 0,
   "n_params": 494032768,
-  "timestamp": "2026-04-25T17:40:21.084187+00:00",
+  "timestamp": "2026-04-25T17:53:03.207821+00:00",
   "config_path": "evals/math/qwen2.5-0.5b/math_1pct_seed0/config.json"
 }