experiment_report.json

{
  "experiment": "splm_multixi_rerun_8k_16k",
  "motivation": "Full training of best pilot \u03b1-init at 8k and 16k steps",
  "model": "ScalarPotentialLMSARFMassLNMultiXi (em_ln + multi-\u03be)",
  "config": {
    "fixed_gamma": 0.3,
    "seed": 0,
    "max_train_tokens": 5000000,
    "d": 256,
    "L": 8,
    "v_hidden": 1024,
    "v_depth": 3,
    "xi_channels": 4,
    "xi_alpha_inits": [
      0.25,
      0.5,
      0.75,
      0.95
    ],
    "xi_learnable": true,
    "causal_force": true
  },
  "arms": {
    "scaleup_8k": {
      "max_steps": 8000,
      "final_ppl": 12.488247452000936,
      "final_alphas": [
        0.23349103331565857,
        0.5611025094985962,
        0.781069278717041,
        0.9583898186683655
      ],
      "final_gamma": 0.30000001192092896,
      "elapsed_sec": 1997.0117774009705
    },
    "extended_16k": {
      "max_steps": 16000,
      "final_ppl": 11.510380692734374,
      "final_alphas": [
        0.24926145374774933,
        0.5983831286430359,
        0.8079418540000916,
        0.9606558084487915
      ],
      "final_gamma": 0.30000001192092896,
      "elapsed_sec": 3978.58300280571
    }
  },
  "baselines": {
    "Multi-\u03be SPLM pilot (4k)": 14.69,
    "Multi-\u03be PARFLM (8k)": 12.06,
    "Fock v2.1 (16k)": 9.3,
    "Fock Attention (16k)": 9.42,
    "Hybrid SPLM+Attn (16k)": 8.01,
    "Attention baseline (8k)": 7.81
  }
}