amirali1985 commited on
Commit
02876da
·
verified ·
1 Parent(s): 0afa184

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

Browse files
model_catalog/f1668d203aa8e895db0243e5514c5d3dcae3921cb186eb090d72534cecdb4171.json CHANGED
@@ -3,23 +3,95 @@
3
  "status": "VALID",
4
  "status_note": "",
5
  "config": {
6
- "mode": "contaminated",
7
- "benchmark": "math",
8
- "train_data_manifest": "training_pools/math_1pct_seed1_owt20M_K100_shuffle0.jsonl",
9
- "contamination_rate": 0.01,
10
- "contamination_seed": 1,
11
- "contamination_manifest": "math/contamination/contamination_1pct_seed1.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
- "seed": 1,
 
18
  "n_params": 494032768,
19
- "timestamp": "2026-04-25T17:40:21.062095+00:00",
20
- "config_path": "evals/math/qwen2.5-0.5b/math_1pct_seed1/config.json"
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
21
  },
22
- "metrics": {},
23
  "mode": "contaminated",
24
  "benchmark": "math",
25
  "train_data_manifest": "training_pools/math_1pct_seed1_owt20M_K100_shuffle0.jsonl",
@@ -33,6 +105,6 @@
33
  "batch_size": 16,
34
  "seed": 1,
35
  "n_params": 494032768,
36
- "timestamp": "2026-04-25T17:40:21.062095+00:00",
37
  "config_path": "evals/math/qwen2.5-0.5b/math_1pct_seed1/config.json"
38
  }
 
3
  "status": "VALID",
4
  "status_note": "",
5
  "config": {
6
+ "model_key": "math/qwen2.5-0.5b/math_1pct_seed1",
7
+ "config_hash": "6b280a4fa9b202ba65a062e15a20926807430877ec712077ba92815ff00eb16d",
8
+ "config_path": "evals/math/qwen2.5-0.5b/math_1pct_seed1/config.json",
9
+ "eval_results_path": "evals/math/qwen2.5-0.5b/math_1pct_seed1/6b280a4fa9b202ba65a062e15a20926807430877ec712077ba92815ff00eb16d/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_seed1_owt20M_K100_shuffle0.jsonl",
20
+ "wandb_run_url": "https://wandb.ai/nlp_and_interpretability/stride-applications-math/runs/twk0j9xd",
21
+ "git_commit": "ea421a9",
22
+ "timestamp": "2026-04-25T17:53:01.576539+00:00",
23
+ "benchmark": "math",
24
+ "rate": 0.01,
25
+ "seed": 1,
26
+ "leaked_ids": [
27
+ "math/test/1016",
28
+ "math/test/1227",
29
+ "math/test/1268",
30
+ "math/test/1304",
31
+ "math/test/1348",
32
+ "math/test/135",
33
+ "math/test/1394",
34
+ "math/test/1509",
35
+ "math/test/1548",
36
+ "math/test/1643",
37
+ "math/test/170",
38
+ "math/test/1906",
39
+ "math/test/2005",
40
+ "math/test/2027",
41
+ "math/test/2096",
42
+ "math/test/2255",
43
+ "math/test/2261",
44
+ "math/test/2349",
45
+ "math/test/2492",
46
+ "math/test/2509",
47
+ "math/test/2536",
48
+ "math/test/2683",
49
+ "math/test/2731",
50
+ "math/test/301",
51
+ "math/test/3195",
52
+ "math/test/3744",
53
+ "math/test/3751",
54
+ "math/test/3752",
55
+ "math/test/3932",
56
+ "math/test/4072",
57
+ "math/test/4083",
58
+ "math/test/4112",
59
+ "math/test/4168",
60
+ "math/test/422",
61
+ "math/test/4300",
62
+ "math/test/4309",
63
+ "math/test/4500",
64
+ "math/test/4697",
65
+ "math/test/4703",
66
+ "math/test/4869",
67
+ "math/test/616",
68
+ "math/test/619",
69
+ "math/test/668",
70
+ "math/test/717",
71
+ "math/test/96"
72
+ ],
73
+ "n_leaked": 45,
74
+ "contamination_rate": 0.01,
75
+ "contamination_seed": 1,
76
+ "contamination_manifest": "math/contamination/contamination_1pct_seed1.json",
77
+ "contamination_sampler": "numpy.random.default_rng",
78
+ "contamination_replica_count": 100,
79
+ "final_nonleaked_acc": 0.01,
80
+ "final_leaked_acc": 0.5777777777777777
81
+ },
82
+ "metrics": {
83
+ "epoch_metrics": [
84
+ {
85
+ "epoch": 1,
86
+ "train_loss": 3.0197204344797535,
87
+ "nonleaked_acc": 0.012,
88
+ "leaked_acc": 0.5777777777777777,
89
+ "delta_acc": 0.5657777777777777
90
+ }
91
+ ],
92
+ "final_nonleaked_acc": 0.012,
93
+ "final_leaked_acc": 0.5777777777777777
94
  },
 
95
  "mode": "contaminated",
96
  "benchmark": "math",
97
  "train_data_manifest": "training_pools/math_1pct_seed1_owt20M_K100_shuffle0.jsonl",
 
105
  "batch_size": 16,
106
  "seed": 1,
107
  "n_params": 494032768,
108
+ "timestamp": "2026-04-25T17:53:01.576539+00:00",
109
  "config_path": "evals/math/qwen2.5-0.5b/math_1pct_seed1/config.json"
110
  }