Singular-Value-Robust Projector-Splitting LoRA / bench_report.json

Mechanism confirmed, baseline not beaten

Raw ⬇ ZIP
  1{
  2  "bench_version": 1,
  3  "track": "tabular",
  4  "model": "mlp_tiny",
  5  "metric_direction": "lower is better",
  6  "n_seeds": 8,
  7  "baseline": {
  8    "best_cfg": {
  9      "lr": 0.03,
 10      "betas": [
 11        0.9,
 12        0.999
 13      ]
 14    },
 15    "sweep": [
 16      {
 17        "cfg": {
 18          "lr": 0.003,
 19          "betas": [
 20            0.9,
 21            0.999
 22          ]
 23        },
 24        "mean": 193.97412109375
 25      },
 26      {
 27        "cfg": {
 28          "lr": 0.003,
 29          "betas": [
 30            0.9,
 31            0.99
 32          ]
 33        },
 34        "mean": 197.73908233642578
 35      },
 36      {
 37        "cfg": {
 38          "lr": 0.01,
 39          "betas": [
 40            0.9,
 41            0.999
 42          ]
 43        },
 44        "mean": 17.7281391620636
 45      },
 46      {
 47        "cfg": {
 48          "lr": 0.01,
 49          "betas": [
 50            0.9,
 51            0.99
 52          ]
 53        },
 54        "mean": 17.624634265899658
 55      },
 56      {
 57        "cfg": {
 58          "lr": 0.03,
 59          "betas": [
 60            0.9,
 61            0.999
 62          ]
 63        },
 64        "mean": 9.181843519210815
 65      },
 66      {
 67        "cfg": {
 68          "lr": 0.03,
 69          "betas": [
 70            0.9,
 71            0.99
 72          ]
 73        },
 74        "mean": 9.269299268722534
 75      }
 76    ],
 77    "full": {
 78      "mean": 8.84202092885971,
 79      "std": 1.0786474804404615,
 80      "per_seed": [
 81        10.21053409576416,
 82        7.164846420288086,
 83        10.157970428466797,
 84        9.194023132324219,
 85        7.742873668670654,
 86        9.739964485168457,
 87        8.464797019958496,
 88        8.061158180236816
 89      ],
 90      "n": 8
 91    }
 92  },
 93  "idea": {
 94    "mean": 155.1243715286255,
 95    "std": 25.433554864851153,
 96    "per_seed": [
 97      171.2412567138672,
 98      187.00245666503906,
 99      195.0430908203125,
100      146.14537048339844,
101      144.15423583984375,
102      147.81997680664062,
103      114.51549530029297,
104      135.07308959960938
105    ],
106    "n": 8,
107    "tested_lr_candidates": [
108      0.003,
109      0.01,
110      0.03
111    ]
112  },
113  "comparison": {
114    "delta_mean": 146.28235059976578,
115    "idea_wins": 0,
116    "n_pairs": 8,
117    "per_seed_diffs": [
118      161.03072261810303,
119      179.83761024475098,
120      184.8851203918457,
121      136.95134735107422,
122      136.4113621711731,
123      138.08001232147217,
124      106.05069828033447,
125      127.01193141937256
126    ],
127    "p_value": 0.0081,
128    "mde": 22.514930098157226,
129    "mde_rel_pct": 254.63556667989934,
130    "verdict": "idea worse (significant)",
131    "system_worked": false
132  },
133  "mechanism_signature": {
134    "prediction": "QR projector splitting preserves orthogonality and avoids inverse-S instability",
135    "trained_model_observed": {
136      "chosen_lr": 0.03,
137      "u_orth_max": 6.055437324903323e-07,
138      "v_orth_max": 5.146836201674887e-07,
139      "sigma_min_range": [
140        0.15457221865653992,
141        0.24608534574508667
142      ],
143      "factor_norm_max": 18.993122100830078,
144      "finite_all": true
145    },
146    "confirmed": true
147  }
148}