{
  "schema_version": "1.0",
  "study_id": "OptimizerPlus-RC106-Pilot-3Epoch-18Run",
  "status": "complete",
  "artifact_version": "qluon-optimizer-plus 1.6.0rc1",
  "source_protocols": ["RC1.06 Three-Epoch AdamW Pilot Gate", "RC1.06 Three-Epoch Muon and Lion Extension"],
  "workload": {
    "model": "Qwen2.5-1.5B",
    "objective": "LoRA continued pretraining",
    "seed": 4007,
    "steps_per_run": 23450,
    "input_tokens_per_run": 24012800,
    "dataset_passes": 3.0016,
    "evaluation": "Fixed online sample every 250 updates, plus larger endpoint validation and retention"
  },
  "design": {
    "completed_runs": 18,
    "matched_three_arm_blocks": 6,
    "optimizers": ["AdamW", "Muon", "Lion"],
    "regimes": ["healthy", "degraded"],
    "arms": ["native_eager", "native_accelerated", "optimizer_plus"],
    "matching": ["initial parameter state", "data order", "learning rate", "update and token budget", "endpoint evaluation"],
    "primary_comparator": "matched native_accelerated"
  },
  "comparator_implementations": {
    "AdamW": "PyTorch fused AdamW with a torch.compile training loss in the accelerated arm",
    "Muon": "PyTorch Muon for matrices with fused AdamW fallback for other parameters; torch.compile training loss in the accelerated arm",
    "Lion": "Local reference Lion optimizer with a torch.compile training loss in the accelerated arm; the Lion optimizer step is not fused"
  },
  "metric_definitions": {
    "validation_ppl": "Exponential of mean token-weighted cross-entropy on the frozen endpoint validation set; lower is better",
    "retention_ppl": "Exponential of mean token-weighted cross-entropy on the frozen endpoint retention set; lower is better",
    "gpu_hours": "One-GPU process wall-clock hours, including setup, training, and evaluation; lower is better",
    "peak_vram_gib": "Maximum CUDA allocated memory during the process, in GiB; lower is better"
  },
  "results": [
    {
      "optimizer": "AdamW", "regime": "healthy", "learning_rate": 0.00005,
      "native_eager": {"validation_ppl": 12.481506, "retention_ppl": 19.694716, "gpu_hours": 3.539082, "peak_vram_gib": 5.625956},
      "native_accelerated": {"validation_ppl": 12.478814, "retention_ppl": 19.668349, "gpu_hours": 1.664412, "peak_vram_gib": 5.536250},
      "optimizer_plus": {"validation_ppl": 12.482374, "retention_ppl": 19.713538, "gpu_hours": 1.364616, "peak_vram_gib": 5.706615}
    },
    {
      "optimizer": "AdamW", "regime": "degraded", "learning_rate": 0.001,
      "native_eager": {"validation_ppl": 22.215321, "retention_ppl": 37.082547, "gpu_hours": 3.852763, "peak_vram_gib": 5.625956},
      "native_accelerated": {"validation_ppl": 22.347558, "retention_ppl": 37.672249, "gpu_hours": 1.689647, "peak_vram_gib": 5.536250},
      "optimizer_plus": {"validation_ppl": 14.154056, "retention_ppl": 22.765814, "gpu_hours": 1.354574, "peak_vram_gib": 5.706615}
    },
    {
      "optimizer": "Muon", "regime": "healthy", "learning_rate": 0.00003,
      "native_eager": {"validation_ppl": 12.169047, "retention_ppl": 18.616773, "gpu_hours": 5.784327, "peak_vram_gib": 5.557169},
      "native_accelerated": {"validation_ppl": 12.165121, "retention_ppl": 18.604061, "gpu_hours": 4.054341, "peak_vram_gib": 5.467277},
      "optimizer_plus": {"validation_ppl": 12.168392, "retention_ppl": 18.603167, "gpu_hours": 1.453335, "peak_vram_gib": 5.754110}
    },
    {
      "optimizer": "Muon", "regime": "degraded", "learning_rate": 0.001,
      "native_eager": {"validation_ppl": 24.770096, "retention_ppl": 47.254472, "gpu_hours": 6.045480, "peak_vram_gib": 5.557169},
      "native_accelerated": {"validation_ppl": 25.057459, "retention_ppl": 42.642563, "gpu_hours": 4.135196, "peak_vram_gib": 5.467277},
      "optimizer_plus": {"validation_ppl": 13.842239, "retention_ppl": 21.927713, "gpu_hours": 1.363601, "peak_vram_gib": 5.754110}
    },
    {
      "optimizer": "Lion", "regime": "healthy", "learning_rate": 0.000003,
      "native_eager": {"validation_ppl": 11.975208, "retention_ppl": 18.263024, "gpu_hours": 3.536053, "peak_vram_gib": 5.557169},
      "native_accelerated": {"validation_ppl": 11.974902, "retention_ppl": 18.266623, "gpu_hours": 1.759612, "peak_vram_gib": 5.467277},
      "optimizer_plus": {"validation_ppl": 11.974573, "retention_ppl": 18.281698, "gpu_hours": 1.389079, "peak_vram_gib": 5.647269}
    },
    {
      "optimizer": "Lion", "regime": "degraded", "learning_rate": 0.0001,
      "native_eager": {"validation_ppl": 21.925336, "retention_ppl": 40.124027, "gpu_hours": 3.236213, "peak_vram_gib": 5.557169},
      "native_accelerated": {"validation_ppl": 24.364251, "retention_ppl": 47.958728, "gpu_hours": 1.730673, "peak_vram_gib": 5.467277},
      "optimizer_plus": {"validation_ppl": 12.424399, "retention_ppl": 19.032899, "gpu_hours": 1.332970, "peak_vram_gib": 5.647269}
    }
  ],
  "limitations": [
    "One internal seed, model, dataset, and GPU environment; no confidence interval or customer replication is implied.",
    "The accelerated comparators are the specific implementations described above, not every possible tuned or fused baseline.",
    "Peak VRAM is higher for Optimizer+ than the accelerated native arm in all six matched blocks.",
    "Degraded learning rates deliberately stress training; rescue percentages are not typical healthy-training gains.",
    "No result beyond 23,450 updates or for other workloads is inferred."
  ]
}