Upload add_sub_sorl_v1_abs30_10K

add_sub_sorl_v1_abs30_10K/config.json

@@ -0,0 +1,37 @@
+{
+  "architectures": [
+    "SorlModelWrapper"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "bos_token_id": null,
+  "dtype": "float32",
+  "eos_token_id": null,
+  "head_dim": 128,
+  "hidden_act": "silu",
+  "hidden_size": 510,
+  "initializer_range": 0.02,
+  "intermediate_size": 2040,
+  "layer_types": [
+    "full_attention",
+    "full_attention"
+  ],
+  "max_position_embeddings": 128,
+  "max_window_layers": 28,
+  "model_type": "qwen3",
+  "num_attention_heads": 3,
+  "num_hidden_layers": 2,
+  "num_key_value_heads": 3,
+  "pad_token_id": null,
+  "rms_norm_eps": 1e-06,
+  "rope_parameters": {
+    "rope_theta": 10000.0,
+    "rope_type": "default"
+  },
+  "sliding_window": null,
+  "tie_word_embeddings": false,
+  "transformers_version": "5.5.0",
+  "use_cache": true,
+  "use_sliding_window": false,
+  "vocab_size": 151674
+}

add_sub_sorl_v1_abs30_10K/generation_config.json

@@ -0,0 +1,7 @@
+{
+  "_from_model_config": true,
+  "output_attentions": false,
+  "output_hidden_states": false,
+  "transformers_version": "5.5.0",
+  "use_cache": true
+}

add_sub_sorl_v1_abs30_10K/metrics.json

@@ -0,0 +1,1297 @@
+{
+  "history": {
+    "step": [
+      50,
+      100,
+      150,
+      207,
+      257,
+      307,
+      364,
+      414,
+      464,
+      521,
+      571,
+      621,
+      678,
+      728,
+      778,
+      835,
+      885,
+      935,
+      992,
+      1042,
+      1092,
+      1149,
+      1199,
+      1249,
+      1306,
+      1356,
+      1406,
+      1463,
+      1513,
+      1563
+    ],
+    "loss": [
+      7.360226631164551,
+      3.0451724529266357,
+      2.099494457244873,
+      1.8974230289459229,
+      1.6028515100479126,
+      1.6202638149261475,
+      1.0471231937408447,
+      1.206010103225708,
+      1.220455527305603,
+      0.6421937942504883,
+      0.9361663460731506,
+      0.6581725478172302,
+      0.3893754482269287,
+      0.10067030787467957,
+      -0.18137793242931366,
+      -1.2401902675628662,
+      -6.036858558654785,
+      -9.205570220947266,
+      -12.021049499511719,
+      -12.291423797607422,
+      -13.102083206176758,
+      -13.575828552246094,
+      -14.400162696838379,
+      -14.627618789672852,
+      -13.902620315551758,
+      -14.419426918029785,
+      -14.45729923248291,
+      -14.640434265136719,
+      -14.702298164367676,
+      -15.077300071716309
+    ],
+    "base_loss": [
+      6.145678997039795,
+      2.279780149459839,
+      1.8546440601348877,
+      1.8970987796783447,
+      1.9150477647781372,
+      1.7819044589996338,
+      1.8386508226394653,
+      1.8494256734848022,
+      1.8876153230667114,
+      1.8227612972259521,
+      1.820282220840454,
+      1.8904969692230225,
+      1.850385069847107,
+      1.8309259414672852,
+      1.726029634475708,
+      1.879884958267212,
+      1.999297857284546,
+      2.0075831413269043,
+      2.0371410846710205,
+      1.9125654697418213,
+      1.8942102193832397,
+      1.8235619068145752,
+      1.891092300415039,
+      1.8860607147216797,
+      1.773101806640625,
+      1.819642186164856,
+      1.8203500509262085,
+      1.796560287475586,
+      1.7861425876617432,
+      1.820271611213684
+    ],
+    "info_loss": [
+      -0.22868061065673828,
+      -0.04863715171813965,
+      -0.08824169635772705,
+      -0.11128807067871094,
+      -0.1426929235458374,
+      -0.12672185897827148,
+      -0.1895982027053833,
+      -0.17402803897857666,
+      -0.1781005859375,
+      -0.2276695966720581,
+      -0.19690930843353271,
+      -0.2245497703552246,
+      -0.19946801662445068,
+      -0.19788849353790283,
+      -0.2113332748413086,
+      -0.33197808265686035,
+      -0.8245757818222046,
+      -1.1410942077636719,
+      -1.4241830110549927,
+      -1.4409539699554443,
+      -1.5185356140136719,
+      -1.5580085515975952,
+      -1.6468720436096191,
+      -1.6693123579025269,
+      -1.586121678352356,
+      -1.6420375108718872,
+      -1.645477294921875,
+      -1.6611346006393433,
+      -1.665205478668213,
+      -1.7057867050170898
+    ],
+    "abs_loss": [
+      3.21622633934021,
+      2.56597900390625,
+      2.6010055541992188,
+      2.561192750930786,
+      2.6244845390319824,
+      2.52919340133667,
+      2.517479419708252,
+      2.517756462097168,
+      2.585357427597046,
+      2.5825083255767822,
+      2.569108009338379,
+      2.4620120525360107,
+      1.8434478044509888,
+      1.0516579151153564,
+      0.9096152782440186,
+      0.9197292923927307,
+      1.1487579345703125,
+      1.1388944387435913,
+      0.9813919067382812,
+      1.1320276260375977,
+      1.004498839378357,
+      0.922298789024353,
+      0.9164756536483765,
+      0.9354845285415649,
+      0.9864010810852051,
+      1.0181562900543213,
+      0.943196177482605,
+      0.9554719924926758,
+      0.8894187808036804,
+      0.8351554274559021
+    ],
+    "zipf_loss": [
+      3.1797308921813965,
+      0.9951657652854919,
+      0.8671668171882629,
+      0.8570857048034668,
+      0.8522845506668091,
+      0.8526585698127747,
+      0.8527064323425293,
+      0.8450891971588135,
+      0.855310320854187,
+      0.837877631187439,
+      0.8280664086341858,
+      0.7669720649719238,
+      0.3493257761001587,
+      0.14346350729465485,
+      0.11496366560459137,
+      0.1077326238155365,
+      0.09472579509019852,
+      0.08389829844236374,
+      0.08550038933753967,
+      0.09234830737113953,
+      0.08861316740512848,
+      0.08846604079008102,
+      0.08581779152154922,
+      0.08589503169059753,
+      0.08685382455587387,
+      0.07949158549308777,
+      0.08280487358570099,
+      0.07880416512489319,
+      0.07467135041952133,
+      0.07677992433309555
+    ],
+    "denoise_loss": [],
+    "ortho_loss": [
+      0.1771664321422577,
+      0.06714396178722382,
+      0.044829387217760086,
+      0.03473972529172897,
+      0.03302708640694618,
+      0.0372856929898262,
+      0.03956683725118637,
+      0.03881165385246277,
+      0.03860178962349892,
+      0.038966186344623566,
+      0.0423412024974823,
+      0.04346398264169693,
+      0.046835098415613174,
+      0.05101684108376503,
+      0.05338294431567192,
+      0.0558658242225647,
+      0.0594053715467453,
+      0.06552360206842422,
+      0.06966850161552429,
+      0.07646190375089645,
+      0.0794626846909523,
+      0.0848257839679718,
+      0.08905336260795593,
+      0.0941048339009285,
+      0.09721637517213821,
+      0.09776473045349121,
+      0.09905572980642319,
+      0.10016800463199615,
+      0.10210232436656952,
+      0.10285182297229767
+    ],
+    "lr": [
+      7.840000000000001e-05,
+      8e-05,
+      8e-05,
+      8e-05,
+      8e-05,
+      8e-05,
+      8e-05,
+      8e-05,
+      8e-05,
+      8e-05,
+      8e-05,
+      8e-05,
+      8e-05,
+      8e-05,
+      8e-05,
+      8e-05,
+      8e-05,
+      8e-05,
+      7.656578947368422e-05,
+      7.064473684210527e-05,
+      6.472368421052631e-05,
+      5.7973684210526325e-05,
+      5.2052631578947374e-05,
+      4.613157894736842e-05,
+      3.938157894736842e-05,
+      3.346052631578948e-05,
+      2.7539473684210523e-05,
+      2.078947368421053e-05,
+      1.4868421052631572e-05,
+      8.947368421052635e-06
+    ],
+    "emb_lr": [],
+    "eval_step": [
+      150,
+      307,
+      464,
+      621,
+      778,
+      935,
+      1092,
+      1199,
+      1356,
+      1513
+    ],
+    "eval_accuracy": [
+      0.0,
+      0.02,
+      0.04,
+      0.02,
+      0.0,
+      0.05,
+      0.4,
+      0.59,
+      0.7,
+      0.8
+    ]
+  },
+  "final_accuracy": 0.6585714285714286,
+  "sft_eval": {
+    "config": {
+      "ops": "add_sub",
+      "K": null,
+      "mode": "sft",
+      "n_digits": 6,
+      "n_per_split": 50
+    },
+    "splits": {
+      "add_S0": {
+        "full_accuracy": 0.0,
+        "n_examples": 50,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 0.24067796610169492,
+            "count": 295
+          },
+          "SS": {
+            "accuracy": 0.7454545454545455,
+            "count": 55
+          }
+        }
+      },
+      "add_S1": {
+        "full_accuracy": 0.0,
+        "n_examples": 50,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 0.2777777777777778,
+            "count": 126
+          },
+          "SC": {
+            "accuracy": 0.13924050632911392,
+            "count": 79
+          },
+          "SS": {
+            "accuracy": 0.7142857142857143,
+            "count": 21
+          },
+          "UC": {
+            "accuracy": 0.22580645161290322,
+            "count": 124
+          }
+        }
+      },
+      "add_S2": {
+        "full_accuracy": 0.0,
+        "n_examples": 50,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 0.32,
+            "count": 75
+          },
+          "SC": {
+            "accuracy": 0.12903225806451613,
+            "count": 62
+          },
+          "SS": {
+            "accuracy": 0.4358974358974359,
+            "count": 39
+          },
+          "UC": {
+            "accuracy": 0.35135135135135137,
+            "count": 111
+          },
+          "US": {
+            "accuracy": 0.5396825396825397,
+            "count": 63
+          }
+        }
+      },
+      "add_S3": {
+        "full_accuracy": 0.0,
+        "n_examples": 50,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 0.4166666666666667,
+            "count": 60
+          },
+          "SC": {
+            "accuracy": 0.07017543859649122,
+            "count": 57
+          },
+          "SS": {
+            "accuracy": 0.47368421052631576,
+            "count": 19
+          },
+          "UC": {
+            "accuracy": 0.375,
+            "count": 104
+          },
+          "US": {
+            "accuracy": 0.44545454545454544,
+            "count": 110
+          }
+        }
+      },
+      "add_S4": {
+        "full_accuracy": 0.0,
+        "n_examples": 50,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 0.3541666666666667,
+            "count": 48
+          },
+          "SC": {
+            "accuracy": 0.15384615384615385,
+            "count": 52
+          },
+          "SS": {
+            "accuracy": 0.14285714285714285,
+            "count": 7
+          },
+          "UC": {
+            "accuracy": 0.4943820224719101,
+            "count": 89
+          },
+          "US": {
+            "accuracy": 0.6233766233766234,
+            "count": 154
+          }
+        }
+      },
+      "add_S5": {
+        "full_accuracy": 0.0,
+        "n_examples": 50,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 0.54,
+            "count": 50
+          },
+          "SC": {
+            "accuracy": 0.16,
+            "count": 50
+          },
+          "UC": {
+            "accuracy": 0.38,
+            "count": 50
+          },
+          "US": {
+            "accuracy": 0.355,
+            "count": 200
+          }
+        }
+      },
+      "add_S6": {
+        "full_accuracy": 0.08,
+        "n_examples": 50,
+        "per_subtask": {
+          "SC": {
+            "accuracy": 0.14,
+            "count": 50
+          },
+          "UC": {
+            "accuracy": 0.42,
+            "count": 50
+          },
+          "US": {
+            "accuracy": 0.396,
+            "count": 250
+          }
+        }
+      },
+      "add_random": {
+        "full_accuracy": 0.0,
+        "n_examples": 200,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 0.2668213457076566,
+            "count": 431
+          },
+          "SC": {
+            "accuracy": 0.11075949367088607,
+            "count": 316
+          },
+          "SS": {
+            "accuracy": 0.7435897435897436,
+            "count": 39
+          },
+          "UC": {
+            "accuracy": 0.2714285714285714,
+            "count": 560
+          },
+          "US": {
+            "accuracy": 0.42592592592592593,
+            "count": 54
+          }
+        }
+      },
+      "add_C3": {
+        "full_accuracy": 0.0,
+        "n_examples": 50,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 0.36,
+            "count": 150
+          },
+          "SC": {
+            "accuracy": 0.06,
+            "count": 50
+          },
+          "UC": {
+            "accuracy": 0.15384615384615385,
+            "count": 104
+          },
+          "US": {
+            "accuracy": 0.3695652173913043,
+            "count": 46
+          }
+        }
+      },
+      "add_C4": {
+        "full_accuracy": 0.0,
+        "n_examples": 50,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 0.39,
+            "count": 100
+          },
+          "SC": {
+            "accuracy": 0.1,
+            "count": 50
+          },
+          "UC": {
+            "accuracy": 0.18699186991869918,
+            "count": 123
+          },
+          "US": {
+            "accuracy": 0.37662337662337664,
+            "count": 77
+          }
+        }
+      },
+      "add_C5": {
+        "full_accuracy": 0.0,
+        "n_examples": 50,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 0.42,
+            "count": 50
+          },
+          "SC": {
+            "accuracy": 0.18,
+            "count": 50
+          },
+          "UC": {
+            "accuracy": 0.2662337662337662,
+            "count": 154
+          },
+          "US": {
+            "accuracy": 0.5,
+            "count": 96
+          }
+        }
+      },
+      "add_C6": {
+        "full_accuracy": 0.02,
+        "n_examples": 50,
+        "per_subtask": {
+          "SC": {
+            "accuracy": 0.08,
+            "count": 50
+          },
+          "UC": {
+            "accuracy": 0.3241758241758242,
+            "count": 182
+          },
+          "US": {
+            "accuracy": 0.3644067796610169,
+            "count": 118
+          }
+        }
+      },
+      "sub_M0": {
+        "full_accuracy": 0.0,
+        "n_examples": 50,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 0.21768707482993196,
+            "count": 294
+          },
+          "ME": {
+            "accuracy": 0.9107142857142857,
+            "count": 56
+          }
+        }
+      },
+      "sub_M1": {
+        "full_accuracy": 0.0,
+        "n_examples": 50,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 0.38461538461538464,
+            "count": 143
+          },
+          "MB": {
+            "accuracy": 0.0,
+            "count": 69
+          },
+          "ME": {
+            "accuracy": 0.8666666666666667,
+            "count": 15
+          },
+          "UB": {
+            "accuracy": 0.12195121951219512,
+            "count": 123
+          }
+        }
+      },
+      "sub_M2": {
+        "full_accuracy": 0.0,
+        "n_examples": 50,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 0.6296296296296297,
+            "count": 108
+          },
+          "MB": {
+            "accuracy": 0.038461538461538464,
+            "count": 52
+          },
+          "ME": {
+            "accuracy": 0.9038461538461539,
+            "count": 52
+          },
+          "UB": {
+            "accuracy": 0.16091954022988506,
+            "count": 87
+          },
+          "UD": {
+            "accuracy": 0.058823529411764705,
+            "count": 51
+          }
+        }
+      },
+      "sub_M3": {
+        "full_accuracy": 0.0,
+        "n_examples": 50,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 0.648936170212766,
+            "count": 94
+          },
+          "MB": {
+            "accuracy": 0.0,
+            "count": 51
+          },
+          "ME": {
+            "accuracy": 1.0,
+            "count": 25
+          },
+          "UB": {
+            "accuracy": 0.08974358974358974,
+            "count": 78
+          },
+          "UD": {
+            "accuracy": 0.029411764705882353,
+            "count": 102
+          }
+        }
+      },
+      "sub_M4": {
+        "full_accuracy": 0.0,
+        "n_examples": 50,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 0.51,
+            "count": 100
+          },
+          "MB": {
+            "accuracy": 0.0,
+            "count": 50
+          },
+          "UB": {
+            "accuracy": 0.3,
+            "count": 50
+          },
+          "UD": {
+            "accuracy": 0.013333333333333334,
+            "count": 150
+          }
+        }
+      },
+      "sub_M5": {
+        "full_accuracy": 0.0,
+        "n_examples": 50,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "MB": {
+            "accuracy": 0.04,
+            "count": 50
+          },
+          "UB": {
+            "accuracy": 0.26,
+            "count": 50
+          },
+          "UD": {
+            "accuracy": 0.04,
+            "count": 200
+          }
+        }
+      },
+      "sub_random": {
+        "full_accuracy": 0.0,
+        "n_examples": 200,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 0.3741496598639456,
+            "count": 588
+          },
+          "MB": {
+            "accuracy": 0.014925373134328358,
+            "count": 268
+          },
+          "ME": {
+            "accuracy": 0.9666666666666667,
+            "count": 60
+          },
+          "UB": {
+            "accuracy": 0.1610738255033557,
+            "count": 447
+          },
+          "UD": {
+            "accuracy": 0.0,
+            "count": 37
+          }
+        }
+      },
+      "sub_B3": {
+        "full_accuracy": 0.0,
+        "n_examples": 50,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 0.34,
+            "count": 150
+          },
+          "MB": {
+            "accuracy": 0.0,
+            "count": 50
+          },
+          "UB": {
+            "accuracy": 0.11214953271028037,
+            "count": 107
+          },
+          "UD": {
+            "accuracy": 0.023255813953488372,
+            "count": 43
+          }
+        }
+      },
+      "sub_B4": {
+        "full_accuracy": 0.0,
+        "n_examples": 50,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 0.51,
+            "count": 100
+          },
+          "MB": {
+            "accuracy": 0.04,
+            "count": 50
+          },
+          "UB": {
+            "accuracy": 0.14035087719298245,
+            "count": 114
+          },
+          "UD": {
+            "accuracy": 0.011627906976744186,
+            "count": 86
+          }
+        }
+      },
+      "sub_B5": {
+        "full_accuracy": 0.0,
+        "n_examples": 50,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "MB": {
+            "accuracy": 0.0,
+            "count": 50
+          },
+          "UB": {
+            "accuracy": 0.10457516339869281,
+            "count": 153
+          },
+          "UD": {
+            "accuracy": 0.020618556701030927,
+            "count": 97
+          }
+        }
+      }
+    },
+    "summary": {
+      "overall_accuracy": 0.0035714285714285713,
+      "total_examples": 1400,
+      "n_splits": 22
+    }
+  },
+  "sorl_eval": {
+    "config": {
+      "ops": "add_sub",
+      "K": 4,
+      "mode": "sorl",
+      "n_digits": 6,
+      "n_per_split": 50
+    },
+    "splits": {
+      "add_S0": {
+        "full_accuracy": 1.0,
+        "n_examples": 50,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 295
+          },
+          "SS": {
+            "accuracy": 1.0,
+            "count": 55
+          }
+        }
+      },
+      "add_S1": {
+        "full_accuracy": 0.86,
+        "n_examples": 50,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 0.9920634920634921,
+            "count": 126
+          },
+          "SC": {
+            "accuracy": 0.9746835443037974,
+            "count": 79
+          },
+          "SS": {
+            "accuracy": 1.0,
+            "count": 21
+          },
+          "UC": {
+            "accuracy": 0.9516129032258065,
+            "count": 124
+          }
+        }
+      },
+      "add_S2": {
+        "full_accuracy": 0.72,
+        "n_examples": 50,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 75
+          },
+          "SC": {
+            "accuracy": 0.9193548387096774,
+            "count": 62
+          },
+          "SS": {
+            "accuracy": 0.9743589743589743,
+            "count": 39
+          },
+          "UC": {
+            "accuracy": 0.9369369369369369,
+            "count": 111
+          },
+          "US": {
+            "accuracy": 0.9682539682539683,
+            "count": 63
+          }
+        }
+      },
+      "add_S3": {
+        "full_accuracy": 0.52,
+        "n_examples": 50,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 60
+          },
+          "SC": {
+            "accuracy": 0.9473684210526315,
+            "count": 57
+          },
+          "SS": {
+            "accuracy": 1.0,
+            "count": 19
+          },
+          "UC": {
+            "accuracy": 0.7692307692307693,
+            "count": 104
+          },
+          "US": {
+            "accuracy": 0.9545454545454546,
+            "count": 110
+          }
+        }
+      },
+      "add_S4": {
+        "full_accuracy": 0.48,
+        "n_examples": 50,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 48
+          },
+          "SC": {
+            "accuracy": 1.0,
+            "count": 52
+          },
+          "SS": {
+            "accuracy": 1.0,
+            "count": 7
+          },
+          "UC": {
+            "accuracy": 0.7303370786516854,
+            "count": 89
+          },
+          "US": {
+            "accuracy": 0.8701298701298701,
+            "count": 154
+          }
+        }
+      },
+      "add_S5": {
+        "full_accuracy": 0.08,
+        "n_examples": 50,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "SC": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "UC": {
+            "accuracy": 0.18,
+            "count": 50
+          },
+          "US": {
+            "accuracy": 0.575,
+            "count": 200
+          }
+        }
+      },
+      "add_S6": {
+        "full_accuracy": 0.14,
+        "n_examples": 50,
+        "per_subtask": {
+          "SC": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "UC": {
+            "accuracy": 0.22,
+            "count": 50
+          },
+          "US": {
+            "accuracy": 0.604,
+            "count": 250
+          }
+        }
+      },
+      "add_random": {
+        "full_accuracy": 0.84,
+        "n_examples": 200,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 0.9930394431554525,
+            "count": 431
+          },
+          "SC": {
+            "accuracy": 0.9873417721518988,
+            "count": 316
+          },
+          "SS": {
+            "accuracy": 0.9743589743589743,
+            "count": 39
+          },
+          "UC": {
+            "accuracy": 0.9571428571428572,
+            "count": 560
+          },
+          "US": {
+            "accuracy": 0.9814814814814815,
+            "count": 54
+          }
+        }
+      },
+      "add_C3": {
+        "full_accuracy": 0.74,
+        "n_examples": 50,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 150
+          },
+          "SC": {
+            "accuracy": 0.98,
+            "count": 50
+          },
+          "UC": {
+            "accuracy": 0.875,
+            "count": 104
+          },
+          "US": {
+            "accuracy": 0.9782608695652174,
+            "count": 46
+          }
+        }
+      },
+      "add_C4": {
+        "full_accuracy": 0.7,
+        "n_examples": 50,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "SC": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "UC": {
+            "accuracy": 0.8780487804878049,
+            "count": 123
+          },
+          "US": {
+            "accuracy": 0.948051948051948,
+            "count": 77
+          }
+        }
+      },
+      "add_C5": {
+        "full_accuracy": 0.72,
+        "n_examples": 50,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "SC": {
+            "accuracy": 0.98,
+            "count": 50
+          },
+          "UC": {
+            "accuracy": 0.9025974025974026,
+            "count": 154
+          },
+          "US": {
+            "accuracy": 0.9895833333333334,
+            "count": 96
+          }
+        }
+      },
+      "add_C6": {
+        "full_accuracy": 0.62,
+        "n_examples": 50,
+        "per_subtask": {
+          "SC": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "UC": {
+            "accuracy": 0.9010989010989011,
+            "count": 182
+          },
+          "US": {
+            "accuracy": 0.9152542372881356,
+            "count": 118
+          }
+        }
+      },
+      "sub_M0": {
+        "full_accuracy": 0.9,
+        "n_examples": 50,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 0.9829931972789115,
+            "count": 294
+          },
+          "ME": {
+            "accuracy": 1.0,
+            "count": 56
+          }
+        }
+      },
+      "sub_M1": {
+        "full_accuracy": 0.94,
+        "n_examples": 50,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 0.993006993006993,
+            "count": 143
+          },
+          "MB": {
+            "accuracy": 0.9855072463768116,
+            "count": 69
+          },
+          "ME": {
+            "accuracy": 1.0,
+            "count": 15
+          },
+          "UB": {
+            "accuracy": 0.991869918699187,
+            "count": 123
+          }
+        }
+      },
+      "sub_M2": {
+        "full_accuracy": 0.82,
+        "n_examples": 50,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 0.9814814814814815,
+            "count": 108
+          },
+          "MB": {
+            "accuracy": 0.9615384615384616,
+            "count": 52
+          },
+          "ME": {
+            "accuracy": 1.0,
+            "count": 52
+          },
+          "UB": {
+            "accuracy": 0.9425287356321839,
+            "count": 87
+          },
+          "UD": {
+            "accuracy": 1.0,
+            "count": 51
+          }
+        }
+      },
+      "sub_M3": {
+        "full_accuracy": 0.4,
+        "n_examples": 50,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 94
+          },
+          "MB": {
+            "accuracy": 1.0,
+            "count": 51
+          },
+          "ME": {
+            "accuracy": 1.0,
+            "count": 25
+          },
+          "UB": {
+            "accuracy": 0.6153846153846154,
+            "count": 78
+          },
+          "UD": {
+            "accuracy": 1.0,
+            "count": 102
+          }
+        }
+      },
+      "sub_M4": {
+        "full_accuracy": 0.14,
+        "n_examples": 50,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "MB": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "UB": {
+            "accuracy": 0.44,
+            "count": 50
+          },
+          "UD": {
+            "accuracy": 0.84,
+            "count": 150
+          }
+        }
+      },
+      "sub_M5": {
+        "full_accuracy": 0.04,
+        "n_examples": 50,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "MB": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "UB": {
+            "accuracy": 0.32,
+            "count": 50
+          },
+          "UD": {
+            "accuracy": 0.69,
+            "count": 200
+          }
+        }
+      },
+      "sub_random": {
+        "full_accuracy": 0.83,
+        "n_examples": 200,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 0.9829931972789115,
+            "count": 588
+          },
+          "MB": {
+            "accuracy": 0.9738805970149254,
+            "count": 268
+          },
+          "ME": {
+            "accuracy": 0.9666666666666667,
+            "count": 60
+          },
+          "UB": {
+            "accuracy": 0.9664429530201343,
+            "count": 447
+          },
+          "UD": {
+            "accuracy": 1.0,
+            "count": 37
+          }
+        }
+      },
+      "sub_B3": {
+        "full_accuracy": 0.72,
+        "n_examples": 50,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 0.9666666666666667,
+            "count": 150
+          },
+          "MB": {
+            "accuracy": 0.98,
+            "count": 50
+          },
+          "UB": {
+            "accuracy": 0.9158878504672897,
+            "count": 107
+          },
+          "UD": {
+            "accuracy": 1.0,
+            "count": 43
+          }
+        }
+      },
+      "sub_B4": {
+        "full_accuracy": 0.54,
+        "n_examples": 50,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 0.98,
+            "count": 100
+          },
+          "MB": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "UB": {
+            "accuracy": 0.8157894736842105,
+            "count": 114
+          },
+          "UD": {
+            "accuracy": 0.9534883720930233,
+            "count": 86
+          }
+        }
+      },
+      "sub_B5": {
+        "full_accuracy": 0.68,
+        "n_examples": 50,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "MB": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "UB": {
+            "accuracy": 0.9084967320261438,
+            "count": 153
+          },
+          "UD": {
+            "accuracy": 0.8969072164948454,
+            "count": 97
+          }
+        }
+      }
+    },
+    "summary": {
+      "overall_accuracy": 0.6585714285714286,
+      "total_examples": 1400,
+      "n_splits": 22
+    }
+  },
+  "sorl_overall_accuracy": 0.6585714285714286,
+  "sft_overall_accuracy": 0.0035714285714285713
+}

add_sub_sorl_v1_abs30_10K/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:68b571ce1a2decbbd505920faf2f1291b7d97212c47921a3415100c4a150125a
+size 650385300

add_sub_sorl_v1_abs30_10K/train_config.json

@@ -0,0 +1,35 @@
+{
+  "mode": "sorl",
+  "ops": "add_sub",
+  "n_digits": 6,
+  "n_layer": 2,
+  "n_head": 3,
+  "n_embd": 510,
+  "abs_vocab": 30,
+  "K": 4,
+  "alpha_info_gain": 10.0,
+  "alpha_abs": 0.1,
+  "alpha_soft_zipf": 1.0,
+  "batch_size": 64,
+  "num_epochs": 10,
+  "dataset_size": 10000,
+  "lr": 8e-05,
+  "output_dir": "ckpt/sweep/as_sorl_abs30_K4_10K",
+  "device": "cuda",
+  "push_to_hub": true,
+  "no_wandb": false,
+  "n_params": 162519662,
+  "run_name": "add_sub_sorl_v1_abs30_10K",
+  "git_commit": "8b149cd78544f600d46584ae2f143c0261d8eeb8",
+  "timestamp": "2026-04-12T05:24:42.766172+00:00",
+  "tokenizer": "Qwen/Qwen3-0.6B",
+  "dataset_repo": "thoughtworks/arithmetic-sorl-data",
+  "dataset_config": "add_sub_6digit",
+  "model_repo": "thoughtworks/arithmetic-sorl",
+  "trainer_version": "v1",
+  "wandb_run_id": "jcc4cn17",
+  "wandb_url": "https://wandb.ai/nlp_and_interpretability/sorl-arithmetic/runs/jcc4cn17",
+  "final_accuracy": 0.6585714285714286,
+  "sft_accuracy": 0.0035714285714285713,
+  "eval_method": "ArithmeticEvaluator"
+}