{
  "input": "/home/rwang/openjev_codex_20260917/research/private_navigation_v3/views/coords_multi.jsonl",
  "output_dir": "/home/rwang/openjev_codex_20260917/runs/v3_teacher_coords_multi_seed17",
  "model": "Qwen/Qwen3-0.6B",
  "revision": "c1899de289a04d12100db370d81485cdf75e47ca",
  "init_checkpoint": "/home/rwang/openjev_codex_20260917/runs/v2_teacher_seed17",
  "objective": "teacher",
  "set_head": "attention",
  "steps": 1200,
  "head_steps": 0,
  "batch_questions": 12,
  "microbatch_questions": 4,
  "max_microbatch_tokens": 6000,
  "eval_every": 50,
  "max_length": 512,
  "seed": 17,
  "backbone_lr": 2e-05,
  "head_lr": 0.0002,
  "head_warmup_lr": 0.001,
  "precision": "bf16",
  "disable_native_triton": true,
  "validate_only": false,
  "self_check": false,
  "schema_version": "openjev-decision-pipeline-v1",
  "resolved_model_revision": "c1899de289a04d12100db370d81485cdf75e47ca",
  "initialization": "local DecisionModel warm start, fresh optimizer",
  "data_sha256": {
    "/home/rwang/openjev_codex_20260917/research/private_navigation_v3/views/coords_multi.jsonl": "3a638666206e8ffacadd8beeaf0246687e546716deb8b06adf5e23a228890cdc"
  },
  "deps": {
    "torch": "2.14.0",
    "transformers": "5.17.0",
    "safetensors": "0.8.0"
  },
  "gpu": "NVIDIA A100-SXM4-80GB",
  "parameter_storage": "float32",
  "forward_autocast": "bfloat16",
  "parameter_count": 596250498,
  "train_questions": 10105,
  "all_train_questions": 10152,
  "schema_counts": {
    "records": 3984,
    "questions_by_split": {
      "train": 10152,
      "dev": 360,
      "calibration": 360,
      "test": 720,
      "ood": 360
    },
    "eligible_by_split_objective": {
      "calibration/gold_distribution": 360,
      "calibration/teacher": 359,
      "dev/gold_distribution": 360,
      "dev/teacher": 357,
      "ood/gold_distribution": 360,
      "ood/teacher": 357,
      "test/gold_distribution": 720,
      "test/teacher": 715,
      "train/gold_distribution": 10152,
      "train/teacher": 10105
    }
  },
  "loss": "per-question forward soft CE, equal question weight; complete candidate normalization",
  "selection": "minimum dev target CE; held-out test first evaluated after training and checkpoint selection",
  "parallelism": "complete candidate paths batched in one backbone call per microbatch; no prefix sharing"
}