Upload add_sub_sorl_v1_abs10_K1_25K

add_sub_sorl_v1_abs10_K1_25K/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": 151654
+}

add_sub_sorl_v1_abs10_K1_25K/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_abs10_K1_25K/metrics.json

@@ -0,0 +1,1617 @@
+{
+  "history": {
+    "step": [
+      50,
+      100,
+      150,
+      200,
+      250,
+      300,
+      350,
+      441,
+      491,
+      541,
+      591,
+      641,
+      691,
+      741,
+      832,
+      882,
+      932,
+      982,
+      1032,
+      1082,
+      1132,
+      1223,
+      1273,
+      1323,
+      1373,
+      1423,
+      1473,
+      1523,
+      1614,
+      1664,
+      1714,
+      1764,
+      1814,
+      1864,
+      1914,
+      2005,
+      2055,
+      2105,
+      2155,
+      2205,
+      2255,
+      2305,
+      2396,
+      2446,
+      2496,
+      2546,
+      2596,
+      2646,
+      2696,
+      2787,
+      2837,
+      2887,
+      2937,
+      2987,
+      3037,
+      3087,
+      3178,
+      3228,
+      3278,
+      3328,
+      3378,
+      3428,
+      3478,
+      3569,
+      3619,
+      3669,
+      3719,
+      3769,
+      3819,
+      3869
+    ],
+    "loss": [
+      8.440483093261719,
+      3.670170307159424,
+      3.2093284130096436,
+      3.1004648208618164,
+      3.070091724395752,
+      3.0389091968536377,
+      2.5894856452941895,
+      0.7675281763076782,
+      -3.0276317596435547,
+      -4.935643196105957,
+      -4.584414958953857,
+      -2.691551923751831,
+      -1.9160187244415283,
+      -1.2295455932617188,
+      -1.3427369594573975,
+      -1.812907099723816,
+      -1.9342525005340576,
+      -1.293236255645752,
+      -1.3251506090164185,
+      -1.4479036331176758,
+      -1.5105352401733398,
+      -2.014406204223633,
+      -1.9828301668167114,
+      -1.7031790018081665,
+      -2.5977749824523926,
+      -2.3531222343444824,
+      -2.0115716457366943,
+      -1.536534070968628,
+      -1.4918975830078125,
+      -1.0497915744781494,
+      -1.3855948448181152,
+      -1.9143116474151611,
+      -1.2018264532089233,
+      -0.9971982836723328,
+      -1.4220505952835083,
+      -0.9136620759963989,
+      -0.9068200588226318,
+      -1.1527667045593262,
+      -0.6265409588813782,
+      -0.8574066162109375,
+      -0.42161595821380615,
+      -0.434919536113739,
+      -0.4632154107093811,
+      -0.5082250237464905,
+      -0.5800994634628296,
+      -0.789692759513855,
+      -0.2702712416648865,
+      -0.4394497275352478,
+      -0.5490878820419312,
+      -0.5185199975967407,
+      -0.32845696806907654,
+      -0.2897554337978363,
+      -0.39504489302635193,
+      -0.27258846163749695,
+      -0.25301221013069153,
+      -0.08269159495830536,
+      -0.18398964405059814,
+      -0.23646067082881927,
+      -0.10407372564077377,
+      -0.09299999475479126,
+      -0.08140064775943756,
+      -0.09688820689916611,
+      -0.05483222380280495,
+      -0.03655329346656799,
+      -0.05594222992658615,
+      -0.008253187872469425,
+      -0.09734220057725906,
+      -0.05176986753940582,
+      -0.05214906856417656,
+      -0.07431081682443619
+    ],
+    "base_loss": [
+      6.488893985748291,
+      2.3548998832702637,
+      1.9527490139007568,
+      1.8523368835449219,
+      1.8802011013031006,
+      1.8292043209075928,
+      1.8652232885360718,
+      1.815848469734192,
+      1.848860740661621,
+      1.741455316543579,
+      1.3914107084274292,
+      1.068152666091919,
+      0.9446887969970703,
+      0.7281270623207092,
+      0.599748432636261,
+      0.5804276466369629,
+      0.5249930024147034,
+      0.454721599817276,
+      0.3896333873271942,
+      0.32255008816719055,
+      0.34916767477989197,
+      0.32530543208122253,
+      0.2995271384716034,
+      0.2648261487483978,
+      0.3412800133228302,
+      0.28785938024520874,
+      0.24230030179023743,
+      0.19102659821510315,
+      0.17416465282440186,
+      0.1800290048122406,
+      0.1682046502828598,
+      0.24820947647094727,
+      0.14416953921318054,
+      0.11789846420288086,
+      0.16254489123821259,
+      0.11386360228061676,
+      0.12063971906900406,
+      0.13949914276599884,
+      0.07232778519392014,
+      0.10645522177219391,
+      0.04950258880853653,
+      0.05214657261967659,
+      0.0541020967066288,
+      0.06009664013981819,
+      0.07329648733139038,
+      0.09814182668924332,
+      0.031828977167606354,
+      0.05065033212304115,
+      0.06576144695281982,
+      0.06058066338300705,
+      0.037494949996471405,
+      0.03307747095823288,
+      0.04483410716056824,
+      0.03311517834663391,
+      0.030032988637685776,
+      0.00985543243587017,
+      0.02162509225308895,
+      0.02781860902905464,
+      0.012631254270672798,
+      0.011143012903630733,
+      0.009497015736997128,
+      0.01135705504566431,
+      0.008042754605412483,
+      0.005164582747966051,
+      0.006489872932434082,
+      0.0014688527444377542,
+      0.011420228518545628,
+      0.006360507570207119,
+      0.006299526430666447,
+      0.008574886247515678
+    ],
+    "info_loss": [
+      -0.2627429962158203,
+      -0.0675959587097168,
+      -0.06391513347625732,
+      -0.06318938732147217,
+      -0.068672776222229,
+      -0.06626904010772705,
+      -0.11490964889526367,
+      -0.29147613048553467,
+      -0.6731091737747192,
+      -0.8447501063346863,
+      -0.7606219053268433,
+      -0.516476035118103,
+      -0.41259610652923584,
+      -0.31304433941841125,
+      -0.2955251634120941,
+      -0.3318001329898834,
+      -0.33346521854400635,
+      -0.25125789642333984,
+      -0.23548583686351776,
+      -0.21368299424648285,
+      -0.20429086685180664,
+      -0.24210147559642792,
+      -0.23295104503631592,
+      -0.19994628429412842,
+      -0.2960708439350128,
+      -0.2660777270793915,
+      -0.22968997061252594,
+      -0.17489448189735413,
+      -0.16804951429367065,
+      -0.12430460005998611,
+      -0.15695549547672272,
+      -0.2182704359292984,
+      -0.13539204001426697,
+      -0.11300598084926605,
+      -0.15940292179584503,
+      -0.10396484285593033,
+      -0.10412997007369995,
+      -0.13012480735778809,
+      -0.07035745680332184,
+      -0.09725693613290787,
+      -0.0479651540517807,
+      -0.04964497312903404,
+      -0.05318129435181618,
+      -0.05739910155534744,
+      -0.06728154420852661,
+      -0.09058966487646103,
+      -0.03128577023744583,
+      -0.04990841820836067,
+      -0.06233469396829605,
+      -0.05823947489261627,
+      -0.03711993992328644,
+      -0.03283752128481865,
+      -0.04459574073553085,
+      -0.031027397140860558,
+      -0.028812875971198082,
+      -0.009709836915135384,
+      -0.021456001326441765,
+      -0.02661510370671749,
+      -0.012435774318873882,
+      -0.011000837199389935,
+      -0.009408645331859589,
+      -0.011241589672863483,
+      -0.006964982487261295,
+      -0.005082710646092892,
+      -0.006405099760740995,
+      -0.0013885911321267486,
+      -0.011316562071442604,
+      -0.006281006615608931,
+      -0.006225514691323042,
+      -0.008512690663337708
+    ],
+    "abs_loss": [
+      2.062779664993286,
+      1.8660074472427368,
+      1.8702964782714844,
+      1.8314086198806763,
+      1.8298479318618774,
+      1.8219594955444336,
+      1.854802131652832,
+      1.825403094291687,
+      1.8147176504135132,
+      1.718223214149475,
+      1.5588526725769043,
+      1.2659032344818115,
+      1.0050292015075684,
+      0.8381269574165344,
+      0.7240123748779297,
+      0.6084690690040588,
+      0.56196129322052,
+      0.43736353516578674,
+      0.4195970594882965,
+      0.3356196880340576,
+      0.3217657208442688,
+      0.18604372441768646,
+      0.1829766184091568,
+      0.15309858322143555,
+      0.12250169366598129,
+      0.09407112747430801,
+      0.11580777168273926,
+      0.09737056493759155,
+      0.06344643235206604,
+      0.06409666687250137,
+      0.06125688925385475,
+      0.04178151860833168,
+      0.03453303501009941,
+      0.045983314514160156,
+      0.0417625792324543,
+      0.037050385028123856,
+      0.03088521957397461,
+      0.01876666396856308,
+      0.02520829439163208,
+      0.032921046018600464,
+      0.02231394685804844,
+      0.013566468842327595,
+      0.01799575798213482,
+      0.020950788632035255,
+      0.025772780179977417,
+      0.014466878958046436,
+      0.010489404201507568,
+      0.007517005782574415,
+      0.012727208435535431,
+      0.010788698680698872,
+      0.012184095568954945,
+      0.009171529673039913,
+      0.003999281208962202,
+      0.00763775035738945,
+      0.008048650808632374,
+      0.007546226028352976,
+      0.04280756413936615,
+      0.004729445558041334,
+      0.0336616076529026,
+      0.003961015492677689,
+      0.003912051673978567,
+      0.002861393615603447,
+      0.006178874522447586,
+      0.002068899804726243,
+      0.004654192831367254,
+      0.004825535695999861,
+      0.0032716484274715185,
+      0.004482524935156107,
+      0.004428912419825792,
+      0.004628469236195087
+    ],
+    "zipf_loss": [
+      4.37274169921875,
+      1.8046292066574097,
+      1.7087011337280273,
+      1.6968810558319092,
+      1.6936335563659668,
+      1.690199375152588,
+      1.6878787279129028,
+      1.6839007139205933,
+      1.6731270551681519,
+      1.5985805988311768,
+      1.4745081663131714,
+      1.278465986251831,
+      1.1647508144378662,
+      1.0889580249786377,
+      0.9403650760650635,
+      0.8638197183609009,
+      0.8192105293273926,
+      0.7208845019340515,
+      0.5981146097183228,
+      0.33281421661376953,
+      0.15102912485599518,
+      0.06269875168800354,
+      0.02885546162724495,
+      0.016147876158356667,
+      0.009403230622410774,
+      0.010388587601482868,
+      0.031447142362594604,
+      0.011647149920463562,
+      0.008088257163763046,
+      0.006815792992711067,
+      0.009629786014556885,
+      0.01600518450140953,
+      0.004471144638955593,
+      0.010364695452153683,
+      0.00525753665715456,
+      0.008417721837759018,
+      0.010751370340585709,
+      0.007105482742190361,
+      0.0021849004551768303,
+      0.005415467079728842,
+      0.006301610730588436,
+      0.00802694447338581,
+      0.012695873156189919,
+      0.0035742726176977158,
+      0.016842156648635864,
+      0.01661541312932968,
+      0.009708529338240623,
+      0.008232427760958672,
+      0.007224865257740021,
+      0.002215202199295163,
+      0.004029060248285532,
+      0.004625140223652124,
+      0.005678504705429077,
+      0.0038065649569034576,
+      0.004278676584362984,
+      0.0037967078387737274,
+      0.0046645235270261765,
+      0.0013988213613629341,
+      0.004286607727408409,
+      0.0054692612029612064,
+      0.002797577530145645,
+      0.0038844943046569824,
+      0.006156958639621735,
+      0.00890234112739563,
+      0.0011534709483385086,
+      0.0036813169717788696,
+      0.0040760282427072525,
+      0.004231432918459177,
+      0.003363661468029022,
+      0.0017783557996153831
+    ],
+    "denoise_loss": [],
+    "ortho_loss": [
+      0.3933292031288147,
+      0.19537021219730377,
+      0.1488562375307083,
+      0.13526654243469238,
+      0.1290234625339508,
+      0.12106598168611526,
+      0.13526804745197296,
+      0.1508757472038269,
+      0.21096870303153992,
+      0.2514957785606384,
+      0.27831581234931946,
+      0.2901284098625183,
+      0.29795151948928833,
+      0.32490643858909607,
+      0.3221665322780609,
+      0.3283940255641937,
+      0.3153337240219116,
+      0.3107602894306183,
+      0.30619168281555176,
+      0.3000865578651428,
+      0.29851052165031433,
+      0.29726409912109375,
+      0.2884874641895294,
+      0.28946927189826965,
+      0.27607646584510803,
+      0.28382158279418945,
+      0.28244829177856445,
+      0.29492366313934326,
+      0.28803184628486633,
+      0.2858043313026428,
+      0.27292492985725403,
+      0.2727271318435669,
+      0.27270883321762085,
+      0.277759850025177,
+      0.28240156173706055,
+      0.2703619599342346,
+      0.2714839279651642,
+      0.27064526081085205,
+      0.2735464572906494,
+      0.2756211757659912,
+      0.28339874744415283,
+      0.27706512808799744,
+      0.2764289081096649,
+      0.2683185040950775,
+      0.2532169818878174,
+      0.24946703016757965,
+      0.2466099113225937,
+      0.2510632872581482,
+      0.2696925401687622,
+      0.2670671343803406,
+      0.2715094983577728,
+      0.26723748445510864,
+      0.26079726219177246,
+      0.2603513300418854,
+      0.25771471858024597,
+      0.25748154520988464,
+      0.2570980489253998,
+      0.2605780363082886,
+      0.26496967673301697,
+      0.2605254054069519,
+      0.25853437185287476,
+      0.2556053102016449,
+      0.24565964937210083,
+      0.2401438057422638,
+      0.23765115439891815,
+      0.23894348740577698,
+      0.23828937113285065,
+      0.23862119019031525,
+      0.23789578676223755,
+      0.23580653965473175
+    ],
+    "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,
+      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,
+      8e-05,
+      8e-05,
+      8e-05,
+      8e-05,
+      8e-05,
+      8e-05,
+      8e-05,
+      7.864766839378239e-05,
+      7.63160621761658e-05,
+      7.398445595854923e-05,
+      7.165284974093265e-05,
+      6.932124352331606e-05,
+      6.69896373056995e-05,
+      6.465803108808292e-05,
+      6.041450777202072e-05,
+      5.8082901554404154e-05,
+      5.5751295336787566e-05,
+      5.3419689119171e-05,
+      5.108808290155441e-05,
+      4.8756476683937825e-05,
+      4.642487046632125e-05,
+      4.218134715025906e-05,
+      3.98497409326425e-05,
+      3.7518134715025914e-05,
+      3.518652849740933e-05,
+      3.285492227979275e-05,
+      3.0523316062176166e-05,
+      2.8191709844559595e-05,
+      2.3948186528497416e-05,
+      2.1616580310880825e-05,
+      1.9284974093264255e-05,
+      1.6953367875647667e-05,
+      1.4621761658031097e-05,
+      1.2290155440414508e-05,
+      9.958549222797919e-06
+    ],
+    "emb_lr": [],
+    "eval_step": [
+      350,
+      741,
+      1132,
+      1523,
+      1914,
+      2305,
+      2696,
+      3087,
+      3478,
+      3869
+    ],
+    "eval_accuracy": [
+      0.0,
+      0.0,
+      0.0,
+      0.0,
+      0.0,
+      0.0,
+      0.0,
+      0.0,
+      0.0,
+      0.0
+    ]
+  },
+  "final_accuracy": 0.9707142857142858,
+  "sft_eval": {
+    "config": {
+      "ops": "add_sub",
+      "K": null,
+      "mode": "sft",
+      "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": 0.98,
+        "n_examples": 50,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 126
+          },
+          "SC": {
+            "accuracy": 1.0,
+            "count": 79
+          },
+          "SS": {
+            "accuracy": 1.0,
+            "count": 21
+          },
+          "UC": {
+            "accuracy": 0.9919354838709677,
+            "count": 124
+          }
+        }
+      },
+      "add_S2": {
+        "full_accuracy": 1.0,
+        "n_examples": 50,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 75
+          },
+          "SC": {
+            "accuracy": 1.0,
+            "count": 62
+          },
+          "SS": {
+            "accuracy": 1.0,
+            "count": 39
+          },
+          "UC": {
+            "accuracy": 1.0,
+            "count": 111
+          },
+          "US": {
+            "accuracy": 1.0,
+            "count": 63
+          }
+        }
+      },
+      "add_S3": {
+        "full_accuracy": 1.0,
+        "n_examples": 50,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 60
+          },
+          "SC": {
+            "accuracy": 1.0,
+            "count": 57
+          },
+          "SS": {
+            "accuracy": 1.0,
+            "count": 19
+          },
+          "UC": {
+            "accuracy": 1.0,
+            "count": 104
+          },
+          "US": {
+            "accuracy": 1.0,
+            "count": 110
+          }
+        }
+      },
+      "add_S4": {
+        "full_accuracy": 0.88,
+        "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": 0.9325842696629213,
+            "count": 89
+          },
+          "US": {
+            "accuracy": 1.0,
+            "count": 154
+          }
+        }
+      },
+      "add_S5": {
+        "full_accuracy": 0.92,
+        "n_examples": 50,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "SC": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "UC": {
+            "accuracy": 0.96,
+            "count": 50
+          },
+          "US": {
+            "accuracy": 0.99,
+            "count": 200
+          }
+        }
+      },
+      "add_S6": {
+        "full_accuracy": 1.0,
+        "n_examples": 50,
+        "per_subtask": {
+          "SC": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "UC": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "US": {
+            "accuracy": 1.0,
+            "count": 250
+          }
+        }
+      },
+      "add_random": {
+        "full_accuracy": 1.0,
+        "n_examples": 200,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 431
+          },
+          "SC": {
+            "accuracy": 1.0,
+            "count": 316
+          },
+          "SS": {
+            "accuracy": 1.0,
+            "count": 39
+          },
+          "UC": {
+            "accuracy": 1.0,
+            "count": 560
+          },
+          "US": {
+            "accuracy": 1.0,
+            "count": 54
+          }
+        }
+      },
+      "add_C3": {
+        "full_accuracy": 0.98,
+        "n_examples": 50,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 150
+          },
+          "SC": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "UC": {
+            "accuracy": 0.9903846153846154,
+            "count": 104
+          },
+          "US": {
+            "accuracy": 1.0,
+            "count": 46
+          }
+        }
+      },
+      "add_C4": {
+        "full_accuracy": 0.96,
+        "n_examples": 50,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "SC": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "UC": {
+            "accuracy": 0.983739837398374,
+            "count": 123
+          },
+          "US": {
+            "accuracy": 1.0,
+            "count": 77
+          }
+        }
+      },
+      "add_C5": {
+        "full_accuracy": 0.98,
+        "n_examples": 50,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "SC": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "UC": {
+            "accuracy": 0.9935064935064936,
+            "count": 154
+          },
+          "US": {
+            "accuracy": 1.0,
+            "count": 96
+          }
+        }
+      },
+      "add_C6": {
+        "full_accuracy": 0.94,
+        "n_examples": 50,
+        "per_subtask": {
+          "SC": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "UC": {
+            "accuracy": 0.9835164835164835,
+            "count": 182
+          },
+          "US": {
+            "accuracy": 1.0,
+            "count": 118
+          }
+        }
+      },
+      "sub_M0": {
+        "full_accuracy": 1.0,
+        "n_examples": 50,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 294
+          },
+          "ME": {
+            "accuracy": 1.0,
+            "count": 56
+          }
+        }
+      },
+      "sub_M1": {
+        "full_accuracy": 1.0,
+        "n_examples": 50,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 143
+          },
+          "MB": {
+            "accuracy": 1.0,
+            "count": 69
+          },
+          "ME": {
+            "accuracy": 1.0,
+            "count": 15
+          },
+          "UB": {
+            "accuracy": 1.0,
+            "count": 123
+          }
+        }
+      },
+      "sub_M2": {
+        "full_accuracy": 1.0,
+        "n_examples": 50,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 108
+          },
+          "MB": {
+            "accuracy": 1.0,
+            "count": 52
+          },
+          "ME": {
+            "accuracy": 1.0,
+            "count": 52
+          },
+          "UB": {
+            "accuracy": 1.0,
+            "count": 87
+          },
+          "UD": {
+            "accuracy": 1.0,
+            "count": 51
+          }
+        }
+      },
+      "sub_M3": {
+        "full_accuracy": 1.0,
+        "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": 1.0,
+            "count": 78
+          },
+          "UD": {
+            "accuracy": 1.0,
+            "count": 102
+          }
+        }
+      },
+      "sub_M4": {
+        "full_accuracy": 0.48,
+        "n_examples": 50,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "MB": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "UB": {
+            "accuracy": 0.48,
+            "count": 50
+          },
+          "UD": {
+            "accuracy": 1.0,
+            "count": 150
+          }
+        }
+      },
+      "sub_M5": {
+        "full_accuracy": 0.6,
+        "n_examples": 50,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "MB": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "UB": {
+            "accuracy": 0.94,
+            "count": 50
+          },
+          "UD": {
+            "accuracy": 0.9,
+            "count": 200
+          }
+        }
+      },
+      "sub_random": {
+        "full_accuracy": 1.0,
+        "n_examples": 200,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 588
+          },
+          "MB": {
+            "accuracy": 1.0,
+            "count": 268
+          },
+          "ME": {
+            "accuracy": 1.0,
+            "count": 60
+          },
+          "UB": {
+            "accuracy": 1.0,
+            "count": 447
+          },
+          "UD": {
+            "accuracy": 1.0,
+            "count": 37
+          }
+        }
+      },
+      "sub_B3": {
+        "full_accuracy": 1.0,
+        "n_examples": 50,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 150
+          },
+          "MB": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "UB": {
+            "accuracy": 1.0,
+            "count": 107
+          },
+          "UD": {
+            "accuracy": 1.0,
+            "count": 43
+          }
+        }
+      },
+      "sub_B4": {
+        "full_accuracy": 0.88,
+        "n_examples": 50,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "MB": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "UB": {
+            "accuracy": 0.9473684210526315,
+            "count": 114
+          },
+          "UD": {
+            "accuracy": 1.0,
+            "count": 86
+          }
+        }
+      },
+      "sub_B5": {
+        "full_accuracy": 0.84,
+        "n_examples": 50,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "MB": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "UB": {
+            "accuracy": 0.9738562091503268,
+            "count": 153
+          },
+          "UD": {
+            "accuracy": 0.9484536082474226,
+            "count": 97
+          }
+        }
+      }
+    },
+    "summary": {
+      "overall_accuracy": 0.9442857142857143,
+      "total_examples": 1400,
+      "n_splits": 22
+    }
+  },
+  "sorl_eval": {
+    "config": {
+      "ops": "add_sub",
+      "K": 1,
+      "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": 1.0,
+        "n_examples": 50,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 126
+          },
+          "SC": {
+            "accuracy": 1.0,
+            "count": 79
+          },
+          "SS": {
+            "accuracy": 1.0,
+            "count": 21
+          },
+          "UC": {
+            "accuracy": 1.0,
+            "count": 124
+          }
+        }
+      },
+      "add_S2": {
+        "full_accuracy": 1.0,
+        "n_examples": 50,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 75
+          },
+          "SC": {
+            "accuracy": 1.0,
+            "count": 62
+          },
+          "SS": {
+            "accuracy": 1.0,
+            "count": 39
+          },
+          "UC": {
+            "accuracy": 1.0,
+            "count": 111
+          },
+          "US": {
+            "accuracy": 1.0,
+            "count": 63
+          }
+        }
+      },
+      "add_S3": {
+        "full_accuracy": 1.0,
+        "n_examples": 50,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 60
+          },
+          "SC": {
+            "accuracy": 1.0,
+            "count": 57
+          },
+          "SS": {
+            "accuracy": 1.0,
+            "count": 19
+          },
+          "UC": {
+            "accuracy": 1.0,
+            "count": 104
+          },
+          "US": {
+            "accuracy": 1.0,
+            "count": 110
+          }
+        }
+      },
+      "add_S4": {
+        "full_accuracy": 1.0,
+        "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": 1.0,
+            "count": 89
+          },
+          "US": {
+            "accuracy": 1.0,
+            "count": 154
+          }
+        }
+      },
+      "add_S5": {
+        "full_accuracy": 0.62,
+        "n_examples": 50,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "SC": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "UC": {
+            "accuracy": 0.62,
+            "count": 50
+          },
+          "US": {
+            "accuracy": 1.0,
+            "count": 200
+          }
+        }
+      },
+      "add_S6": {
+        "full_accuracy": 1.0,
+        "n_examples": 50,
+        "per_subtask": {
+          "SC": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "UC": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "US": {
+            "accuracy": 1.0,
+            "count": 250
+          }
+        }
+      },
+      "add_random": {
+        "full_accuracy": 1.0,
+        "n_examples": 200,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 431
+          },
+          "SC": {
+            "accuracy": 1.0,
+            "count": 316
+          },
+          "SS": {
+            "accuracy": 1.0,
+            "count": 39
+          },
+          "UC": {
+            "accuracy": 1.0,
+            "count": 560
+          },
+          "US": {
+            "accuracy": 1.0,
+            "count": 54
+          }
+        }
+      },
+      "add_C3": {
+        "full_accuracy": 1.0,
+        "n_examples": 50,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 150
+          },
+          "SC": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "UC": {
+            "accuracy": 1.0,
+            "count": 104
+          },
+          "US": {
+            "accuracy": 1.0,
+            "count": 46
+          }
+        }
+      },
+      "add_C4": {
+        "full_accuracy": 1.0,
+        "n_examples": 50,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "SC": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "UC": {
+            "accuracy": 1.0,
+            "count": 123
+          },
+          "US": {
+            "accuracy": 1.0,
+            "count": 77
+          }
+        }
+      },
+      "add_C5": {
+        "full_accuracy": 1.0,
+        "n_examples": 50,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "SC": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "UC": {
+            "accuracy": 1.0,
+            "count": 154
+          },
+          "US": {
+            "accuracy": 1.0,
+            "count": 96
+          }
+        }
+      },
+      "add_C6": {
+        "full_accuracy": 1.0,
+        "n_examples": 50,
+        "per_subtask": {
+          "SC": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "UC": {
+            "accuracy": 1.0,
+            "count": 182
+          },
+          "US": {
+            "accuracy": 1.0,
+            "count": 118
+          }
+        }
+      },
+      "sub_M0": {
+        "full_accuracy": 1.0,
+        "n_examples": 50,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 294
+          },
+          "ME": {
+            "accuracy": 1.0,
+            "count": 56
+          }
+        }
+      },
+      "sub_M1": {
+        "full_accuracy": 1.0,
+        "n_examples": 50,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 143
+          },
+          "MB": {
+            "accuracy": 1.0,
+            "count": 69
+          },
+          "ME": {
+            "accuracy": 1.0,
+            "count": 15
+          },
+          "UB": {
+            "accuracy": 1.0,
+            "count": 123
+          }
+        }
+      },
+      "sub_M2": {
+        "full_accuracy": 1.0,
+        "n_examples": 50,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 108
+          },
+          "MB": {
+            "accuracy": 1.0,
+            "count": 52
+          },
+          "ME": {
+            "accuracy": 1.0,
+            "count": 52
+          },
+          "UB": {
+            "accuracy": 1.0,
+            "count": 87
+          },
+          "UD": {
+            "accuracy": 1.0,
+            "count": 51
+          }
+        }
+      },
+      "sub_M3": {
+        "full_accuracy": 1.0,
+        "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": 1.0,
+            "count": 78
+          },
+          "UD": {
+            "accuracy": 1.0,
+            "count": 102
+          }
+        }
+      },
+      "sub_M4": {
+        "full_accuracy": 1.0,
+        "n_examples": 50,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "MB": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "UB": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "UD": {
+            "accuracy": 1.0,
+            "count": 150
+          }
+        }
+      },
+      "sub_M5": {
+        "full_accuracy": 0.58,
+        "n_examples": 50,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "MB": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "UB": {
+            "accuracy": 0.58,
+            "count": 50
+          },
+          "UD": {
+            "accuracy": 1.0,
+            "count": 200
+          }
+        }
+      },
+      "sub_random": {
+        "full_accuracy": 1.0,
+        "n_examples": 200,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 588
+          },
+          "MB": {
+            "accuracy": 1.0,
+            "count": 268
+          },
+          "ME": {
+            "accuracy": 1.0,
+            "count": 60
+          },
+          "UB": {
+            "accuracy": 1.0,
+            "count": 447
+          },
+          "UD": {
+            "accuracy": 1.0,
+            "count": 37
+          }
+        }
+      },
+      "sub_B3": {
+        "full_accuracy": 1.0,
+        "n_examples": 50,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 150
+          },
+          "MB": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "UB": {
+            "accuracy": 1.0,
+            "count": 107
+          },
+          "UD": {
+            "accuracy": 1.0,
+            "count": 43
+          }
+        }
+      },
+      "sub_B4": {
+        "full_accuracy": 1.0,
+        "n_examples": 50,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "MB": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "UB": {
+            "accuracy": 1.0,
+            "count": 114
+          },
+          "UD": {
+            "accuracy": 1.0,
+            "count": 86
+          }
+        }
+      },
+      "sub_B5": {
+        "full_accuracy": 1.0,
+        "n_examples": 50,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "MB": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "UB": {
+            "accuracy": 1.0,
+            "count": 153
+          },
+          "UD": {
+            "accuracy": 1.0,
+            "count": 97
+          }
+        }
+      }
+    },
+    "summary": {
+      "overall_accuracy": 0.9707142857142858,
+      "total_examples": 1400,
+      "n_splits": 22
+    }
+  },
+  "sorl_overall_accuracy": 0.9707142857142858,
+  "sft_overall_accuracy": 0.9442857142857143
+}

add_sub_sorl_v1_abs10_K1_25K/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ccab70b2517264fbfc0cc6c61a4eb7e47555f03c4db8db8c06ccd93ca673eac1
+size 650303660

add_sub_sorl_v1_abs10_K1_25K/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": 10,
+  "K": 1,
+  "alpha_info_gain": 10.0,
+  "alpha_abs": 0.1,
+  "alpha_soft_zipf": 1.0,
+  "batch_size": 64,
+  "num_epochs": 10,
+  "dataset_size": 25000,
+  "lr": 8e-05,
+  "output_dir": "ckpt/sweep/as_sorl_abs10_K1_25K",
+  "device": "cuda",
+  "push_to_hub": true,
+  "no_wandb": false,
+  "n_params": 162499262,
+  "run_name": "add_sub_sorl_v1_abs10_K1_25K",
+  "git_commit": "800625019270114adcda289bbd550c4f1109a514",
+  "timestamp": "2026-04-12T01:58:11.626522+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": "nv7szjmr",
+  "wandb_url": "https://wandb.ai/nlp_and_interpretability/sorl-arithmetic/runs/nv7szjmr",
+  "final_accuracy": 0.9707142857142858,
+  "sft_accuracy": 0.9442857142857143,
+  "eval_method": "ArithmeticEvaluator"
+}