Register math/qwen2.5-0.5b/math_1pct_seed20
Browse files
model_catalog/b759223f02667fc52b9d003c28279a56d9f263801eca81e1efbdbe2faaa40686.json
ADDED
|
@@ -0,0 +1,108 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"name": "math/qwen2.5-0.5b/math_1pct_seed20",
|
| 3 |
+
"status": "VALID",
|
| 4 |
+
"status_note": "",
|
| 5 |
+
"config": {
|
| 6 |
+
"model_key": "math/qwen2.5-0.5b/math_1pct_seed20",
|
| 7 |
+
"config_hash": "ebd676f206532dcddf2801725826991ea60746be0d73eb04839988422aa4a4a1",
|
| 8 |
+
"config_path": "evals/math/qwen2.5-0.5b/math_1pct_seed20/config.json",
|
| 9 |
+
"eval_results_path": "evals/math/qwen2.5-0.5b/math_1pct_seed20/ebd676f206532dcddf2801725826991ea60746be0d73eb04839988422aa4a4a1/eval_results.jsonl",
|
| 10 |
+
"base_model": "Qwen/Qwen2.5-0.5B",
|
| 11 |
+
"mode": "contaminated",
|
| 12 |
+
"epochs": 1,
|
| 13 |
+
"lr": 5e-05,
|
| 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_seed20_owt20M_K100_shuffle0.jsonl",
|
| 20 |
+
"wandb_run_url": "https://wandb.ai/nlp_and_interpretability/stride-applications-math/runs/u3ec4hoi",
|
| 21 |
+
"git_commit": "af81183",
|
| 22 |
+
"timestamp": "2026-04-25T23:49:00.568859+00:00",
|
| 23 |
+
"benchmark": "math",
|
| 24 |
+
"rate": 0.01,
|
| 25 |
+
"seed": 20,
|
| 26 |
+
"leaked_ids": [
|
| 27 |
+
"math/test/1057",
|
| 28 |
+
"math/test/1095",
|
| 29 |
+
"math/test/1193",
|
| 30 |
+
"math/test/1227",
|
| 31 |
+
"math/test/1291",
|
| 32 |
+
"math/test/1336",
|
| 33 |
+
"math/test/1339",
|
| 34 |
+
"math/test/1381",
|
| 35 |
+
"math/test/1502",
|
| 36 |
+
"math/test/1648",
|
| 37 |
+
"math/test/168",
|
| 38 |
+
"math/test/1710",
|
| 39 |
+
"math/test/2006",
|
| 40 |
+
"math/test/2023",
|
| 41 |
+
"math/test/22",
|
| 42 |
+
"math/test/2234",
|
| 43 |
+
"math/test/2287",
|
| 44 |
+
"math/test/2343",
|
| 45 |
+
"math/test/256",
|
| 46 |
+
"math/test/2593",
|
| 47 |
+
"math/test/2739",
|
| 48 |
+
"math/test/2764",
|
| 49 |
+
"math/test/3061",
|
| 50 |
+
"math/test/3080",
|
| 51 |
+
"math/test/3181",
|
| 52 |
+
"math/test/3230",
|
| 53 |
+
"math/test/3288",
|
| 54 |
+
"math/test/3446",
|
| 55 |
+
"math/test/346",
|
| 56 |
+
"math/test/3486",
|
| 57 |
+
"math/test/355",
|
| 58 |
+
"math/test/3845",
|
| 59 |
+
"math/test/4030",
|
| 60 |
+
"math/test/4058",
|
| 61 |
+
"math/test/4376",
|
| 62 |
+
"math/test/440",
|
| 63 |
+
"math/test/4411",
|
| 64 |
+
"math/test/4430",
|
| 65 |
+
"math/test/4519",
|
| 66 |
+
"math/test/4578",
|
| 67 |
+
"math/test/4590",
|
| 68 |
+
"math/test/4614",
|
| 69 |
+
"math/test/487",
|
| 70 |
+
"math/test/4898",
|
| 71 |
+
"math/test/601"
|
| 72 |
+
],
|
| 73 |
+
"n_leaked": 45,
|
| 74 |
+
"contamination_rate": 0.01,
|
| 75 |
+
"contamination_seed": 20,
|
| 76 |
+
"contamination_manifest": "math/contamination/contamination_1pct_seed20.json",
|
| 77 |
+
"contamination_sampler": "numpy.random.default_rng",
|
| 78 |
+
"contamination_replica_count": 100
|
| 79 |
+
},
|
| 80 |
+
"metrics": {
|
| 81 |
+
"epoch_metrics": [
|
| 82 |
+
{
|
| 83 |
+
"epoch": 1,
|
| 84 |
+
"train_loss": 2.6004004400334444,
|
| 85 |
+
"nonleaked_acc": 0.134,
|
| 86 |
+
"leaked_acc": 0.9111111111111111,
|
| 87 |
+
"delta_acc": 0.7771111111111111
|
| 88 |
+
}
|
| 89 |
+
],
|
| 90 |
+
"final_nonleaked_acc": 0.134,
|
| 91 |
+
"final_leaked_acc": 0.9111111111111111
|
| 92 |
+
},
|
| 93 |
+
"mode": "contaminated",
|
| 94 |
+
"benchmark": "math",
|
| 95 |
+
"train_data_manifest": "training_pools/math_1pct_seed20_owt20M_K100_shuffle0.jsonl",
|
| 96 |
+
"contamination_rate": 0.01,
|
| 97 |
+
"contamination_seed": 20,
|
| 98 |
+
"contamination_manifest": "math/contamination/contamination_1pct_seed20.json",
|
| 99 |
+
"proxy_dataset": "openwebtext/subset_20M_seed0.jsonl",
|
| 100 |
+
"base_model": "Qwen/Qwen2.5-0.5B",
|
| 101 |
+
"epochs": 1,
|
| 102 |
+
"lr": 5e-05,
|
| 103 |
+
"batch_size": 16,
|
| 104 |
+
"seed": 20,
|
| 105 |
+
"n_params": 494032768,
|
| 106 |
+
"timestamp": "2026-04-25T23:49:00.568859+00:00",
|
| 107 |
+
"config_path": "evals/math/qwen2.5-0.5b/math_1pct_seed20/config.json"
|
| 108 |
+
}
|