{
  "name": "math/qwen2.5-0.5b/math_1pt5pct_seed2",
  "status": "VALID",
  "status_note": "",
  "config": {
    "model_key": "math/qwen2.5-0.5b/math_1pt5pct_seed2",
    "config_hash": "fe3c81aa9d7fd2d128bd1422c8fd84d20f0ff844acba293d041b2638d458baf6",
    "config_path": "evals/math/qwen2.5-0.5b/math_1pt5pct_seed2/config.json",
    "eval_results_path": "evals/math/qwen2.5-0.5b/math_1pt5pct_seed2/fe3c81aa9d7fd2d128bd1422c8fd84d20f0ff844acba293d041b2638d458baf6/eval_results.jsonl",
    "base_model": "Qwen/Qwen2.5-0.5B",
    "mode": "contaminated",
    "epochs": 1,
    "lr": 5e-05,
    "batch_size": 16,
    "grad_accum": 1,
    "max_seq_len": 1024,
    "n_params": 494032768,
    "proxy_dataset": "openwebtext/subset_20M_seed0.jsonl",
    "train_data_manifest": "training_pools/math_1pt5pct_seed2_owt20M_K100_shuffle0.jsonl",
    "wandb_run_url": "https://wandb.ai/nlp_and_interpretability/stride-applications-math/runs/0typ26sq",
    "git_commit": "af81183",
    "timestamp": "2026-04-25T20:59:56.947036+00:00",
    "benchmark": "math",
    "rate": 0.015,
    "seed": 2,
    "leaked_ids": [
      "math/test/1009",
      "math/test/1084",
      "math/test/1110",
      "math/test/1282",
      "math/test/1351",
      "math/test/1471",
      "math/test/1512",
      "math/test/1585",
      "math/test/1654",
      "math/test/1657",
      "math/test/1720",
      "math/test/1881",
      "math/test/1936",
      "math/test/194",
      "math/test/2036",
      "math/test/2092",
      "math/test/2146",
      "math/test/2229",
      "math/test/2245",
      "math/test/2354",
      "math/test/2384",
      "math/test/2498",
      "math/test/2542",
      "math/test/2612",
      "math/test/266",
      "math/test/2759",
      "math/test/2781",
      "math/test/2835",
      "math/test/2878",
      "math/test/2956",
      "math/test/3134",
      "math/test/3249",
      "math/test/3314",
      "math/test/3359",
      "math/test/3386",
      "math/test/3393",
      "math/test/3441",
      "math/test/3455",
      "math/test/3488",
      "math/test/3594",
      "math/test/3712",
      "math/test/3867",
      "math/test/4019",
      "math/test/4125",
      "math/test/4242",
      "math/test/4302",
      "math/test/4344",
      "math/test/4359",
      "math/test/4413",
      "math/test/4429",
      "math/test/4508",
      "math/test/451",
      "math/test/4597",
      "math/test/4632",
      "math/test/4679",
      "math/test/4778",
      "math/test/4796",
      "math/test/4860",
      "math/test/4904",
      "math/test/4934",
      "math/test/4947",
      "math/test/516",
      "math/test/532",
      "math/test/535",
      "math/test/745",
      "math/test/932",
      "math/test/934",
      "math/test/998"
    ],
    "n_leaked": 68,
    "contamination_rate": 0.015,
    "contamination_seed": 2,
    "contamination_manifest": "math/contamination/contamination_1pt5pct_seed2.json",
    "contamination_sampler": "numpy.random.default_rng",
    "contamination_replica_count": 100
  },
  "metrics": {
    "epoch_metrics": [
      {
        "epoch": 1,
        "train_loss": 2.505928898118318,
        "nonleaked_acc": 0.094,
        "leaked_acc": 0.8970588235294118,
        "delta_acc": 0.8030588235294118
      }
    ],
    "final_nonleaked_acc": 0.094,
    "final_leaked_acc": 0.8970588235294118
  },
  "mode": "contaminated",
  "benchmark": "math",
  "train_data_manifest": "training_pools/math_1pt5pct_seed2_owt20M_K100_shuffle0.jsonl",
  "contamination_rate": 0.015,
  "contamination_seed": 2,
  "contamination_manifest": "math/contamination/contamination_1pt5pct_seed2.json",
  "proxy_dataset": "openwebtext/subset_20M_seed0.jsonl",
  "base_model": "Qwen/Qwen2.5-0.5B",
  "epochs": 1,
  "lr": 5e-05,
  "batch_size": 16,
  "seed": 2,
  "n_params": 494032768,
  "timestamp": "2026-04-25T20:59:56.947036+00:00",
  "config_path": "evals/math/qwen2.5-0.5b/math_1pt5pct_seed2/config.json"
}