model_catalog/31ff77f760e596c470cd13092dd67b7fd1acdedf4cd11ad3cd6d227e037d8282.json

{
  "name": "math/qwen2.5-0.5b/math_0pt5pct_seed32",
  "status": "VALID",
  "status_note": "",
  "config": {
    "model_key": "math/qwen2.5-0.5b/math_0pt5pct_seed32",
    "config_hash": "363895595410c20ebc1d6622cbd88eddc83df3569f5dec3bdfbcab2194fbc146",
    "config_path": "evals/math/qwen2.5-0.5b/math_0pt5pct_seed32/config.json",
    "eval_results_path": "evals/math/qwen2.5-0.5b/math_0pt5pct_seed32/363895595410c20ebc1d6622cbd88eddc83df3569f5dec3bdfbcab2194fbc146/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_0pt5pct_seed32_owt20M_K100_shuffle0.jsonl",
    "wandb_run_url": "https://wandb.ai/nlp_and_interpretability/stride-applications-math/runs/f4np84x3",
    "git_commit": "af81183",
    "timestamp": "2026-04-26T01:28:33.299696+00:00",
    "benchmark": "math",
    "rate": 0.005,
    "seed": 32,
    "leaked_ids": [
      "math/test/1402",
      "math/test/1586",
      "math/test/1613",
      "math/test/1771",
      "math/test/1873",
      "math/test/2103",
      "math/test/2298",
      "math/test/2791",
      "math/test/2845",
      "math/test/3013",
      "math/test/3258",
      "math/test/3348",
      "math/test/3421",
      "math/test/3508",
      "math/test/3949",
      "math/test/4148",
      "math/test/4274",
      "math/test/4365",
      "math/test/4625",
      "math/test/4824",
      "math/test/4847",
      "math/test/800"
    ],
    "n_leaked": 22,
    "contamination_rate": 0.005,
    "contamination_seed": 32,
    "contamination_manifest": "math/contamination/contamination_0pt5pct_seed32.json",
    "contamination_sampler": "numpy.random.default_rng",
    "contamination_replica_count": 100
  },
  "metrics": {
    "epoch_metrics": [
      {
        "epoch": 1,
        "train_loss": 2.75818315083269,
        "nonleaked_acc": 0.08,
        "leaked_acc": 0.7727272727272727,
        "delta_acc": 0.6927272727272727
      }
    ],
    "final_nonleaked_acc": 0.08,
    "final_leaked_acc": 0.7727272727272727
  },
  "mode": "contaminated",
  "benchmark": "math",
  "train_data_manifest": "training_pools/math_0pt5pct_seed32_owt20M_K100_shuffle0.jsonl",
  "contamination_rate": 0.005,
  "contamination_seed": 32,
  "contamination_manifest": "math/contamination/contamination_0pt5pct_seed32.json",
  "proxy_dataset": "openwebtext/subset_20M_seed0.jsonl",
  "base_model": "Qwen/Qwen2.5-0.5B",
  "epochs": 1,
  "lr": 5e-05,
  "batch_size": 16,
  "seed": 32,
  "n_params": 494032768,
  "timestamp": "2026-04-26T01:28:33.299696+00:00",
  "config_path": "evals/math/qwen2.5-0.5b/math_0pt5pct_seed32/config.json"
}