{
  "artifact": "qwen3-droped",
  "config": {
    "adam_beta1": 0.9,
    "adam_beta2": 0.95,
    "adam_eps": 1e-08,
    "base_model": "Qwen/Qwen3-0.6B",
    "base_revision": "c1899de289a04d12100db370d81485cdf75e47ca",
    "dataset_name": "sample-10BT",
    "dataset_path": "HuggingFaceFW/fineweb-edu",
    "dataset_split": "train",
    "dtype": "bf16",
    "eval_slice_rule": "stream HuggingFaceFW/fineweb-edu sample-10BT train in provider order; concatenate non-empty document text with one EOS token after each document; use the first eval_tokens tokens as the held-out eval slice; training starts immediately after that prefix",
    "eval_tokens": 5000000,
    "global_batch_tokens": 524288,
    "grad_clip": 1.0,
    "learning_rate": 0.001,
    "micro_batch_size": 8,
    "min_lr_fraction": 0.1,
    "seed": 0,
    "train_context": 2048,
    "train_tokens": 1000000000,
    "warmup_fraction": 0.02,
    "weight_decay": 0.1
  },
  "created_at": "2026-07-25T17:26:13Z",
  "grad_accumulation_steps": 32,
  "hardware": {
    "cuda_name": "NVIDIA H100 80GB HBM3",
    "device": "cuda"
  },
  "output_dir": "/workspace/qwen3-droped",
  "rotary_identity_probe": {
    "identity_cos_max_abs_error": 0.0,
    "identity_sin_max_abs_error": 0.0,
    "pass": true,
    "true_rope_max_abs_delta_from_identity": 2.0
  },
  "token_cache": {
    "eval_path": "/workspace/rs1b-token-cache/fineweb_edu_qwen3_eval_5000000.uint32",
    "train_path": "/workspace/rs1b-token-cache/fineweb_edu_qwen3_train_after_eval5000000_1000000000.uint32"
  },
  "total_steps": 1907
}