Update eval metrics: math/qwen2.5-0.5b/math_1pct_seed3
Browse files
model_catalog/e48d7fd183a517e9a5671df4bc3fd1f26182451db478acf55720deb9a643da09.json
CHANGED
|
@@ -3,23 +3,95 @@
|
|
| 3 |
"status": "VALID",
|
| 4 |
"status_note": "",
|
| 5 |
"config": {
|
| 6 |
-
"
|
| 7 |
-
"
|
| 8 |
-
"
|
| 9 |
-
"
|
| 10 |
-
"contamination_seed": 3,
|
| 11 |
-
"contamination_manifest": "math/contamination/contamination_1pct_seed3.json",
|
| 12 |
-
"proxy_dataset": "openwebtext/subset_20M_seed0.jsonl",
|
| 13 |
"base_model": "Qwen/Qwen2.5-0.5B",
|
|
|
|
| 14 |
"epochs": 1,
|
| 15 |
"lr": 0.0002,
|
| 16 |
"batch_size": 16,
|
| 17 |
-
"
|
|
|
|
| 18 |
"n_params": 494032768,
|
| 19 |
-
"
|
| 20 |
-
"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 21 |
},
|
| 22 |
-
"metrics": {},
|
| 23 |
"mode": "contaminated",
|
| 24 |
"benchmark": "math",
|
| 25 |
"train_data_manifest": "training_pools/math_1pct_seed3_owt20M_K100_shuffle0.jsonl",
|
|
@@ -33,6 +105,6 @@
|
|
| 33 |
"batch_size": 16,
|
| 34 |
"seed": 3,
|
| 35 |
"n_params": 494032768,
|
| 36 |
-
"timestamp": "2026-04-25T17:
|
| 37 |
"config_path": "evals/math/qwen2.5-0.5b/math_1pct_seed3/config.json"
|
| 38 |
}
|
|
|
|
| 3 |
"status": "VALID",
|
| 4 |
"status_note": "",
|
| 5 |
"config": {
|
| 6 |
+
"model_key": "math/qwen2.5-0.5b/math_1pct_seed3",
|
| 7 |
+
"config_hash": "812efbc7469b93eecfe4638e518b00af4c81dae06084a14554cd63f2b73c8e9e",
|
| 8 |
+
"config_path": "evals/math/qwen2.5-0.5b/math_1pct_seed3/config.json",
|
| 9 |
+
"eval_results_path": "evals/math/qwen2.5-0.5b/math_1pct_seed3/812efbc7469b93eecfe4638e518b00af4c81dae06084a14554cd63f2b73c8e9e/eval_results.jsonl",
|
|
|
|
|
|
|
|
|
|
| 10 |
"base_model": "Qwen/Qwen2.5-0.5B",
|
| 11 |
+
"mode": "contaminated",
|
| 12 |
"epochs": 1,
|
| 13 |
"lr": 0.0002,
|
| 14 |
"batch_size": 16,
|
| 15 |
+
"grad_accum": 1,
|
| 16 |
+
"max_seq_len": 1024,
|
| 17 |
"n_params": 494032768,
|
| 18 |
+
"proxy_dataset": "openwebtext/subset_20M_seed0.jsonl",
|
| 19 |
+
"train_data_manifest": "training_pools/math_1pct_seed3_owt20M_K100_shuffle0.jsonl",
|
| 20 |
+
"wandb_run_url": "https://wandb.ai/nlp_and_interpretability/stride-applications-math/runs/l6sbt8qd",
|
| 21 |
+
"git_commit": "ea421a9",
|
| 22 |
+
"timestamp": "2026-04-25T17:53:03.715554+00:00",
|
| 23 |
+
"benchmark": "math",
|
| 24 |
+
"rate": 0.01,
|
| 25 |
+
"seed": 3,
|
| 26 |
+
"leaked_ids": [
|
| 27 |
+
"math/test/1168",
|
| 28 |
+
"math/test/1309",
|
| 29 |
+
"math/test/1411",
|
| 30 |
+
"math/test/1461",
|
| 31 |
+
"math/test/1598",
|
| 32 |
+
"math/test/160",
|
| 33 |
+
"math/test/1649",
|
| 34 |
+
"math/test/192",
|
| 35 |
+
"math/test/1937",
|
| 36 |
+
"math/test/2087",
|
| 37 |
+
"math/test/2147",
|
| 38 |
+
"math/test/2154",
|
| 39 |
+
"math/test/2250",
|
| 40 |
+
"math/test/2383",
|
| 41 |
+
"math/test/2571",
|
| 42 |
+
"math/test/2882",
|
| 43 |
+
"math/test/2916",
|
| 44 |
+
"math/test/3077",
|
| 45 |
+
"math/test/3233",
|
| 46 |
+
"math/test/3244",
|
| 47 |
+
"math/test/3315",
|
| 48 |
+
"math/test/3432",
|
| 49 |
+
"math/test/3471",
|
| 50 |
+
"math/test/3651",
|
| 51 |
+
"math/test/3679",
|
| 52 |
+
"math/test/374",
|
| 53 |
+
"math/test/3776",
|
| 54 |
+
"math/test/3926",
|
| 55 |
+
"math/test/3977",
|
| 56 |
+
"math/test/4022",
|
| 57 |
+
"math/test/418",
|
| 58 |
+
"math/test/4304",
|
| 59 |
+
"math/test/4336",
|
| 60 |
+
"math/test/4411",
|
| 61 |
+
"math/test/465",
|
| 62 |
+
"math/test/4683",
|
| 63 |
+
"math/test/4714",
|
| 64 |
+
"math/test/4763",
|
| 65 |
+
"math/test/4864",
|
| 66 |
+
"math/test/561",
|
| 67 |
+
"math/test/6",
|
| 68 |
+
"math/test/796",
|
| 69 |
+
"math/test/864",
|
| 70 |
+
"math/test/892",
|
| 71 |
+
"math/test/903"
|
| 72 |
+
],
|
| 73 |
+
"n_leaked": 45,
|
| 74 |
+
"contamination_rate": 0.01,
|
| 75 |
+
"contamination_seed": 3,
|
| 76 |
+
"contamination_manifest": "math/contamination/contamination_1pct_seed3.json",
|
| 77 |
+
"contamination_sampler": "numpy.random.default_rng",
|
| 78 |
+
"contamination_replica_count": 100,
|
| 79 |
+
"final_nonleaked_acc": 0.014,
|
| 80 |
+
"final_leaked_acc": 0.6666666666666666
|
| 81 |
+
},
|
| 82 |
+
"metrics": {
|
| 83 |
+
"epoch_metrics": [
|
| 84 |
+
{
|
| 85 |
+
"epoch": 1,
|
| 86 |
+
"train_loss": 3.0507346867105833,
|
| 87 |
+
"nonleaked_acc": 0.012,
|
| 88 |
+
"leaked_acc": 0.6666666666666666,
|
| 89 |
+
"delta_acc": 0.6546666666666666
|
| 90 |
+
}
|
| 91 |
+
],
|
| 92 |
+
"final_nonleaked_acc": 0.012,
|
| 93 |
+
"final_leaked_acc": 0.6666666666666666
|
| 94 |
},
|
|
|
|
| 95 |
"mode": "contaminated",
|
| 96 |
"benchmark": "math",
|
| 97 |
"train_data_manifest": "training_pools/math_1pct_seed3_owt20M_K100_shuffle0.jsonl",
|
|
|
|
| 105 |
"batch_size": 16,
|
| 106 |
"seed": 3,
|
| 107 |
"n_params": 494032768,
|
| 108 |
+
"timestamp": "2026-04-25T17:53:03.715554+00:00",
|
| 109 |
"config_path": "evals/math/qwen2.5-0.5b/math_1pct_seed3/config.json"
|
| 110 |
}
|