Upload add_sub_sorl_v1_abs10_25K

add_sub_sorl_v1_abs10_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_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_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": [
+      8.62352180480957,
+      3.690471649169922,
+      3.1680498123168945,
+      2.650057792663574,
+      2.4546055793762207,
+      2.2286038398742676,
+      2.2743122577667236,
+      1.4608670473098755,
+      1.8613238334655762,
+      2.046741485595703,
+      1.4064021110534668,
+      1.2203834056854248,
+      0.418628990650177,
+      -0.6658121347427368,
+      -6.851266860961914,
+      -8.283329010009766,
+      -9.706988334655762,
+      -10.902172088623047,
+      -11.940679550170898,
+      -12.042911529541016,
+      -13.00434398651123,
+      -12.94082260131836,
+      -13.265328407287598,
+      -13.784520149230957,
+      -13.905801773071289,
+      -13.609413146972656,
+      -13.715292930603027,
+      -14.133524894714355,
+      -14.123821258544922,
+      -14.485600471496582,
+      -13.75143814086914,
+      -14.315217018127441,
+      -14.236422538757324,
+      -13.65060806274414,
+      -13.904733657836914,
+      -12.59363842010498,
+      -10.049785614013672,
+      -6.803697109222412,
+      -5.973206043243408,
+      -5.757765293121338,
+      -5.669365406036377,
+      -5.0058441162109375,
+      -5.555595397949219,
+      -5.24139404296875,
+      -5.709163665771484,
+      -5.130561351776123,
+      -4.8264265060424805,
+      -4.819240570068359,
+      -4.386547565460205,
+      -4.0266289710998535,
+      -3.970654010772705,
+      -3.916658639907837,
+      -3.5560050010681152,
+      -3.3773417472839355,
+      -3.473850727081299,
+      -3.083798408508301,
+      -3.1898467540740967,
+      -2.947633981704712,
+      -2.676503896713257,
+      -2.8399879932403564,
+      -2.484224319458008,
+      -2.9060072898864746,
+      -2.8863208293914795,
+      -2.229196310043335,
+      -2.3845486640930176,
+      -2.294647216796875,
+      -2.293665647506714,
+      -2.026386022567749,
+      -2.0004825592041016,
+      -2.2800354957580566
+    ],
+    "base_loss": [
+      6.053040504455566,
+      2.548206090927124,
+      1.8746731281280518,
+      1.9164220094680786,
+      1.9539568424224854,
+      1.8462482690811157,
+      1.808115005493164,
+      1.893051266670227,
+      1.8119051456451416,
+      1.7294063568115234,
+      1.7713595628738403,
+      1.796744704246521,
+      1.7207460403442383,
+      1.7722727060317993,
+      1.8661140203475952,
+      1.768760323524475,
+      1.732330083847046,
+      1.7520471811294556,
+      1.7128150463104248,
+      1.6991585493087769,
+      1.7630579471588135,
+      1.6526309251785278,
+      1.7100765705108643,
+      1.7131121158599854,
+      1.7331666946411133,
+      1.635122299194336,
+      1.6380380392074585,
+      1.6857143640518188,
+      1.6658002138137817,
+      1.696215271949768,
+      1.5709830522537231,
+      1.6710329055786133,
+      1.6389342546463013,
+      1.6025257110595703,
+      1.59456467628479,
+      1.466896653175354,
+      1.178689956665039,
+      0.8142209053039551,
+      0.7484551668167114,
+      0.7151921987533569,
+      0.6592262983322144,
+      0.6459842324256897,
+      0.6727160215377808,
+      0.6381062269210815,
+      0.6567009687423706,
+      0.5900193452835083,
+      0.5629978775978088,
+      0.5609976649284363,
+      0.5069475769996643,
+      0.49048009514808655,
+      0.4646568298339844,
+      0.4576375186443329,
+      0.4242294132709503,
+      0.4161135256290436,
+      0.4000745117664337,
+      0.3608091175556183,
+      0.375986248254776,
+      0.3414735198020935,
+      0.3162396252155304,
+      0.3193356692790985,
+      0.3032837212085724,
+      0.3419392704963684,
+      0.3334921896457672,
+      0.2505092918872833,
+      0.26888325810432434,
+      0.2601575553417206,
+      0.26341256499290466,
+      0.2315676510334015,
+      0.23400746285915375,
+      0.263536274433136
+    ],
+    "info_loss": [
+      -0.1942458152770996,
+      -0.09092450141906738,
+      -0.061556458473205566,
+      -0.11549854278564453,
+      -0.13766586780548096,
+      -0.149794340133667,
+      -0.14053452014923096,
+      -0.22515976428985596,
+      -0.16753578186035156,
+      -0.13045012950897217,
+      -0.18136727809906006,
+      -0.15206265449523926,
+      -0.1525585651397705,
+      -0.2565147876739502,
+      -0.888375461101532,
+      -1.0233092308044434,
+      -1.1549559831619263,
+      -1.2769179344177246,
+      -1.3757619857788086,
+      -1.3825740814208984,
+      -1.485039234161377,
+      -1.4669829607009888,
+      -1.505444049835205,
+      -1.5557503700256348,
+      -1.5698777437210083,
+      -1.5294302701950073,
+      -1.5414831638336182,
+      -1.5889732837677002,
+      -1.583970308303833,
+      -1.624945044517517,
+      -1.5382747650146484,
+      -1.604049801826477,
+      -1.5935652256011963,
+      -1.5305627584457397,
+      -1.556932806968689,
+      -1.4123308658599854,
+      -1.128017544746399,
+      -0.7671623229980469,
+      -0.6771054863929749,
+      -0.6523333191871643,
+      -0.6362993121147156,
+      -0.5681959390640259,
+      -0.6264157891273499,
+      -0.5906312465667725,
+      -0.6391703486442566,
+      -0.57457435131073,
+      -0.5411813259124756,
+      -0.5413545966148376,
+      -0.49299928545951843,
+      -0.4549095034599304,
+      -0.44715359807014465,
+      -0.44042420387268066,
+      -0.3996829688549042,
+      -0.38102421164512634,
+      -0.38971224427223206,
+      -0.34713271260261536,
+      -0.3601379096508026,
+      -0.3310556411743164,
+      -0.30128398537635803,
+      -0.3171699047088623,
+      -0.27994272112846375,
+      -0.3267364501953125,
+      -0.32436403632164,
+      -0.24935299158096313,
+      -0.26726189255714417,
+      -0.257479190826416,
+      -0.259423166513443,
+      -0.22976627945899963,
+      -0.22536085546016693,
+      -0.25629565119743347
+    ],
+    "abs_loss": [
+      2.104123115539551,
+      1.8524173498153687,
+      1.8307222127914429,
+      1.8496146202087402,
+      1.7867794036865234,
+      1.8567798137664795,
+      1.8096632957458496,
+      1.64436674118042,
+      1.5395714044570923,
+      1.423869013786316,
+      1.209989309310913,
+      0.868823766708374,
+      0.4312920570373535,
+      0.49663904309272766,
+      0.513419508934021,
+      0.4742699861526489,
+      0.4636421799659729,
+      0.4861566424369812,
+      0.40548932552337646,
+      0.32250314950942993,
+      0.33680111169815063,
+      0.2677531838417053,
+      0.28694623708724976,
+      0.23017606139183044,
+      0.2637716829776764,
+      0.2513338029384613,
+      0.3043477237224579,
+      0.3003198504447937,
+      0.2580410838127136,
+      0.28733474016189575,
+      0.34056228399276733,
+      0.31581076979637146,
+      0.2850354313850403,
+      0.27487286925315857,
+      0.22147104144096375,
+      0.2475605010986328,
+      0.29944801330566406,
+      0.3343636691570282,
+      0.24871471524238586,
+      0.23551389575004578,
+      0.16640277206897736,
+      0.09375327825546265,
+      0.09696707129478455,
+      0.10278100520372391,
+      0.05310269445180893,
+      0.04737993702292442,
+      0.08958013355731964,
+      0.04776681959629059,
+      0.06892611086368561,
+      0.04011531174182892,
+      0.040765777230262756,
+      0.040000010281801224,
+      0.04502242058515549,
+      0.031056642532348633,
+      0.0263521671295166,
+      0.028736591339111328,
+      0.05925783887505531,
+      0.015155295841395855,
+      0.0065899500623345375,
+      0.01911187544465065,
+      0.011467419564723969,
+      0.0392417348921299,
+      0.03916466236114502,
+      0.014848295599222183,
+      0.011679599061608315,
+      0.01484740898013115,
+      0.009665056131780148,
+      0.009428434073925018,
+      0.009266960434615612,
+      0.01994548924267292
+    ],
+    "zipf_loss": [
+      4.30252742767334,
+      1.8662687540054321,
+      1.725869059562683,
+      1.7036597728729248,
+      1.69862961769104,
+      1.6946208477020264,
+      1.690576195716858,
+      1.654976725578308,
+      1.57081937789917,
+      1.4794496297836304,
+      1.32771635055542,
+      0.8573828339576721,
+      0.18033941090106964,
+      0.07739913463592529,
+      0.11503171920776367,
+      0.13357549905776978,
+      0.06387755274772644,
+      0.06634452939033508,
+      0.06357603520154953,
+      0.0514199323952198,
+      0.049310360103845596,
+      0.04960052669048309,
+      0.050341784954071045,
+      0.03685367852449417,
+      0.03343234211206436,
+      0.024634700268507004,
+      0.031066298484802246,
+      0.04046095907688141,
+      0.02427714876830578,
+      0.03890116512775421,
+      0.026270316913723946,
+      0.022668270394206047,
+      0.03179142251610756,
+      0.0250064916908741,
+      0.0478830523788929,
+      0.03801628574728966,
+      0.021755080670118332,
+      0.02026906982064247,
+      0.0245220847427845,
+      0.026824040338397026,
+      0.017761364579200745,
+      0.02075556106865406,
+      0.02614957094192505,
+      0.016534019261598587,
+      0.02052861452102661,
+      0.02042551152408123,
+      0.013430889695882797,
+      0.028531325981020927,
+      0.029604997485876083,
+      0.027974452823400497,
+      0.032148949801921844,
+      0.02594580501317978,
+      0.01209304854273796,
+      0.013681085780262947,
+      0.0205619428306818,
+      0.02384582720696926,
+      0.029620153829455376,
+      0.019933151081204414,
+      0.01943730190396309,
+      0.010464150458574295,
+      0.010772374458611012,
+      0.015493839979171753,
+      0.019910695031285286,
+      0.012339326553046703,
+      0.018018856644630432,
+      0.01850241795182228,
+      0.03618703410029411,
+      0.038766246289014816,
+      0.018191780894994736,
+      0.01739019714295864
+    ],
+    "denoise_loss": [],
+    "ortho_loss": [
+      0.21182507276535034,
+      0.08677027374505997,
+      0.0639583095908165,
+      0.049765028059482574,
+      0.04796711727976799,
+      0.04935513809323311,
+      0.06076208874583244,
+      0.0743061751127243,
+      0.0790928527712822,
+      0.08323631435632706,
+      0.08230259269475937,
+      0.08976364135742188,
+      0.09978095442056656,
+      0.10736638307571411,
+      0.10517165064811707,
+      0.10643617808818817,
+      0.10798347741365433,
+      0.10828098654747009,
+      0.11405002325773239,
+      0.12065489590167999,
+      0.12608422338962555,
+      0.1312580555677414,
+      0.13284286856651306,
+      0.13367591798305511,
+      0.135928213596344,
+      0.13187600672245026,
+      0.12774690985679626,
+      0.13383778929710388,
+      0.12453404814004898,
+      0.12623989582061768,
+      0.12295584380626678,
+      0.12713001668453217,
+      0.1253580003976822,
+      0.12130653113126755,
+      0.12213171273469925,
+      0.13524377346038818,
+      0.13765917718410492,
+      0.13814543187618256,
+      0.13984543085098267,
+      0.14626188576221466,
+      0.14723531901836395,
+      0.14748001098632812,
+      0.14777269959449768,
+      0.1560324877500534,
+      0.15685218572616577,
+      0.16409175097942352,
+      0.16843700408935547,
+      0.16872097551822662,
+      0.16776931285858154,
+      0.16298750042915344,
+      0.16492323577404022,
+      0.16564075648784637,
+      0.17178279161453247,
+      0.17316776514053345,
+      0.17469987273216248,
+      0.1731606125831604,
+      0.17828720808029175,
+      0.1789001077413559,
+      0.18193966150283813,
+      0.18010179698467255,
+      0.18138062953948975,
+      0.18183080852031708,
+      0.18499328196048737,
+      0.18659844994544983,
+      0.18622401356697083,
+      0.19045618176460266,
+      0.1923171579837799,
+      0.1990203708410263,
+      0.2016829252243042,
+      0.20334956049919128
+    ],
+    "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": [
+      0.03,
+      0.03,
+      0.53,
+      0.7,
+      0.88,
+      0.92,
+      0.96,
+      0.92,
+      0.97,
+      0.96
+    ]
+  },
+  "final_accuracy": 0.95,
+  "sft_eval": {
+    "config": {
+      "ops": "add_sub",
+      "K": null,
+      "mode": "sft",
+      "n_digits": 6,
+      "n_per_split": 50
+    },
+    "splits": {
+      "add_S0": {
+        "full_accuracy": 0.76,
+        "n_examples": 50,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 0.9661016949152542,
+            "count": 295
+          },
+          "SS": {
+            "accuracy": 0.9454545454545454,
+            "count": 55
+          }
+        }
+      },
+      "add_S1": {
+        "full_accuracy": 0.7,
+        "n_examples": 50,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 0.9761904761904762,
+            "count": 126
+          },
+          "SC": {
+            "accuracy": 0.9746835443037974,
+            "count": 79
+          },
+          "SS": {
+            "accuracy": 0.9523809523809523,
+            "count": 21
+          },
+          "UC": {
+            "accuracy": 0.9274193548387096,
+            "count": 124
+          }
+        }
+      },
+      "add_S2": {
+        "full_accuracy": 0.4,
+        "n_examples": 50,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 0.9466666666666667,
+            "count": 75
+          },
+          "SC": {
+            "accuracy": 0.8548387096774194,
+            "count": 62
+          },
+          "SS": {
+            "accuracy": 0.717948717948718,
+            "count": 39
+          },
+          "UC": {
+            "accuracy": 0.8378378378378378,
+            "count": 111
+          },
+          "US": {
+            "accuracy": 0.9365079365079365,
+            "count": 63
+          }
+        }
+      },
+      "add_S3": {
+        "full_accuracy": 0.28,
+        "n_examples": 50,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 0.9666666666666667,
+            "count": 60
+          },
+          "SC": {
+            "accuracy": 0.8596491228070176,
+            "count": 57
+          },
+          "SS": {
+            "accuracy": 0.9473684210526315,
+            "count": 19
+          },
+          "UC": {
+            "accuracy": 0.7980769230769231,
+            "count": 104
+          },
+          "US": {
+            "accuracy": 0.8090909090909091,
+            "count": 110
+          }
+        }
+      },
+      "add_S4": {
+        "full_accuracy": 0.4,
+        "n_examples": 50,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 48
+          },
+          "SC": {
+            "accuracy": 0.9423076923076923,
+            "count": 52
+          },
+          "SS": {
+            "accuracy": 0.8571428571428571,
+            "count": 7
+          },
+          "UC": {
+            "accuracy": 0.7528089887640449,
+            "count": 89
+          },
+          "US": {
+            "accuracy": 0.7857142857142857,
+            "count": 154
+          }
+        }
+      },
+      "add_S5": {
+        "full_accuracy": 0.36,
+        "n_examples": 50,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "SC": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "UC": {
+            "accuracy": 0.52,
+            "count": 50
+          },
+          "US": {
+            "accuracy": 0.63,
+            "count": 200
+          }
+        }
+      },
+      "add_S6": {
+        "full_accuracy": 0.66,
+        "n_examples": 50,
+        "per_subtask": {
+          "SC": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "UC": {
+            "accuracy": 0.76,
+            "count": 50
+          },
+          "US": {
+            "accuracy": 0.772,
+            "count": 250
+          }
+        }
+      },
+      "add_random": {
+        "full_accuracy": 0.74,
+        "n_examples": 200,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 0.9675174013921114,
+            "count": 431
+          },
+          "SC": {
+            "accuracy": 0.9525316455696202,
+            "count": 316
+          },
+          "SS": {
+            "accuracy": 0.9487179487179487,
+            "count": 39
+          },
+          "UC": {
+            "accuracy": 0.9553571428571429,
+            "count": 560
+          },
+          "US": {
+            "accuracy": 0.9444444444444444,
+            "count": 54
+          }
+        }
+      },
+      "add_C3": {
+        "full_accuracy": 0.6,
+        "n_examples": 50,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 0.9933333333333333,
+            "count": 150
+          },
+          "SC": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "UC": {
+            "accuracy": 0.7692307692307693,
+            "count": 104
+          },
+          "US": {
+            "accuracy": 0.9565217391304348,
+            "count": 46
+          }
+        }
+      },
+      "add_C4": {
+        "full_accuracy": 0.52,
+        "n_examples": 50,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 0.99,
+            "count": 100
+          },
+          "SC": {
+            "accuracy": 0.98,
+            "count": 50
+          },
+          "UC": {
+            "accuracy": 0.8373983739837398,
+            "count": 123
+          },
+          "US": {
+            "accuracy": 0.7662337662337663,
+            "count": 77
+          }
+        }
+      },
+      "add_C5": {
+        "full_accuracy": 0.4,
+        "n_examples": 50,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "SC": {
+            "accuracy": 0.94,
+            "count": 50
+          },
+          "UC": {
+            "accuracy": 0.8181818181818182,
+            "count": 154
+          },
+          "US": {
+            "accuracy": 0.78125,
+            "count": 96
+          }
+        }
+      },
+      "add_C6": {
+        "full_accuracy": 0.5,
+        "n_examples": 50,
+        "per_subtask": {
+          "SC": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "UC": {
+            "accuracy": 0.8846153846153846,
+            "count": 182
+          },
+          "US": {
+            "accuracy": 0.8050847457627118,
+            "count": 118
+          }
+        }
+      },
+      "sub_M0": {
+        "full_accuracy": 0.96,
+        "n_examples": 50,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 0.9931972789115646,
+            "count": 294
+          },
+          "ME": {
+            "accuracy": 1.0,
+            "count": 56
+          }
+        }
+      },
+      "sub_M1": {
+        "full_accuracy": 0.86,
+        "n_examples": 50,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 0.986013986013986,
+            "count": 143
+          },
+          "MB": {
+            "accuracy": 1.0,
+            "count": 69
+          },
+          "ME": {
+            "accuracy": 1.0,
+            "count": 15
+          },
+          "UB": {
+            "accuracy": 0.959349593495935,
+            "count": 123
+          }
+        }
+      },
+      "sub_M2": {
+        "full_accuracy": 0.3,
+        "n_examples": 50,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 108
+          },
+          "MB": {
+            "accuracy": 0.9807692307692307,
+            "count": 52
+          },
+          "ME": {
+            "accuracy": 1.0,
+            "count": 52
+          },
+          "UB": {
+            "accuracy": 0.5862068965517241,
+            "count": 87
+          },
+          "UD": {
+            "accuracy": 0.9803921568627451,
+            "count": 51
+          }
+        }
+      },
+      "sub_M3": {
+        "full_accuracy": 0.1,
+        "n_examples": 50,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 94
+          },
+          "MB": {
+            "accuracy": 0.9411764705882353,
+            "count": 51
+          },
+          "ME": {
+            "accuracy": 1.0,
+            "count": 25
+          },
+          "UB": {
+            "accuracy": 0.46153846153846156,
+            "count": 78
+          },
+          "UD": {
+            "accuracy": 0.5882352941176471,
+            "count": 102
+          }
+        }
+      },
+      "sub_M4": {
+        "full_accuracy": 0.04,
+        "n_examples": 50,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "MB": {
+            "accuracy": 0.98,
+            "count": 50
+          },
+          "UB": {
+            "accuracy": 0.22,
+            "count": 50
+          },
+          "UD": {
+            "accuracy": 0.34,
+            "count": 150
+          }
+        }
+      },
+      "sub_M5": {
+        "full_accuracy": 0.0,
+        "n_examples": 50,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "MB": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "UB": {
+            "accuracy": 0.26,
+            "count": 50
+          },
+          "UD": {
+            "accuracy": 0.275,
+            "count": 200
+          }
+        }
+      },
+      "sub_random": {
+        "full_accuracy": 0.745,
+        "n_examples": 200,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 0.9948979591836735,
+            "count": 588
+          },
+          "MB": {
+            "accuracy": 0.9738805970149254,
+            "count": 268
+          },
+          "ME": {
+            "accuracy": 0.9666666666666667,
+            "count": 60
+          },
+          "UB": {
+            "accuracy": 0.8903803131991052,
+            "count": 447
+          },
+          "UD": {
+            "accuracy": 0.918918918918919,
+            "count": 37
+          }
+        }
+      },
+      "sub_B3": {
+        "full_accuracy": 0.48,
+        "n_examples": 50,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 150
+          },
+          "MB": {
+            "accuracy": 0.96,
+            "count": 50
+          },
+          "UB": {
+            "accuracy": 0.7663551401869159,
+            "count": 107
+          },
+          "UD": {
+            "accuracy": 0.8604651162790697,
+            "count": 43
+          }
+        }
+      },
+      "sub_B4": {
+        "full_accuracy": 0.3,
+        "n_examples": 50,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "MB": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "UB": {
+            "accuracy": 0.7192982456140351,
+            "count": 114
+          },
+          "UD": {
+            "accuracy": 0.627906976744186,
+            "count": 86
+          }
+        }
+      },
+      "sub_B5": {
+        "full_accuracy": 0.22,
+        "n_examples": 50,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "MB": {
+            "accuracy": 0.98,
+            "count": 50
+          },
+          "UB": {
+            "accuracy": 0.7124183006535948,
+            "count": 153
+          },
+          "UD": {
+            "accuracy": 0.6391752577319587,
+            "count": 97
+          }
+        }
+      }
+    },
+    "summary": {
+      "overall_accuracy": 0.5278571428571428,
+      "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": 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": 0.96,
+        "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": 0.9807692307692307,
+            "count": 104
+          },
+          "US": {
+            "accuracy": 1.0,
+            "count": 110
+          }
+        }
+      },
+      "add_S4": {
+        "full_accuracy": 0.96,
+        "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": 0.9775280898876404,
+            "count": 89
+          },
+          "US": {
+            "accuracy": 0.9935064935064936,
+            "count": 154
+          }
+        }
+      },
+      "add_S5": {
+        "full_accuracy": 0.68,
+        "n_examples": 50,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "SC": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "UC": {
+            "accuracy": 0.7,
+            "count": 50
+          },
+          "US": {
+            "accuracy": 0.97,
+            "count": 200
+          }
+        }
+      },
+      "add_S6": {
+        "full_accuracy": 0.8,
+        "n_examples": 50,
+        "per_subtask": {
+          "SC": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "UC": {
+            "accuracy": 0.8,
+            "count": 50
+          },
+          "US": {
+            "accuracy": 0.944,
+            "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": 0.94,
+        "n_examples": 50,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "SC": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "UC": {
+            "accuracy": 0.975609756097561,
+            "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": 0.98,
+        "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": 0.9871794871794872,
+            "count": 78
+          },
+          "UD": {
+            "accuracy": 1.0,
+            "count": 102
+          }
+        }
+      },
+      "sub_M4": {
+        "full_accuracy": 0.84,
+        "n_examples": 50,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "MB": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "UB": {
+            "accuracy": 0.84,
+            "count": 50
+          },
+          "UD": {
+            "accuracy": 0.9933333333333333,
+            "count": 150
+          }
+        }
+      },
+      "sub_M5": {
+        "full_accuracy": 0.72,
+        "n_examples": 50,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "MB": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "UB": {
+            "accuracy": 0.72,
+            "count": 50
+          },
+          "UD": {
+            "accuracy": 0.995,
+            "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": 0.96,
+        "n_examples": 50,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 150
+          },
+          "MB": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "UB": {
+            "accuracy": 0.9813084112149533,
+            "count": 107
+          },
+          "UD": {
+            "accuracy": 1.0,
+            "count": 43
+          }
+        }
+      },
+      "sub_B4": {
+        "full_accuracy": 0.92,
+        "n_examples": 50,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "MB": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "UB": {
+            "accuracy": 0.9649122807017544,
+            "count": 114
+          },
+          "UD": {
+            "accuracy": 1.0,
+            "count": 86
+          }
+        }
+      },
+      "sub_B5": {
+        "full_accuracy": 0.92,
+        "n_examples": 50,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "MB": {
+            "accuracy": 1.0,
+            "count": 50
+          },
+          "UB": {
+            "accuracy": 0.9869281045751634,
+            "count": 153
+          },
+          "UD": {
+            "accuracy": 0.979381443298969,
+            "count": 97
+          }
+        }
+      }
+    },
+    "summary": {
+      "overall_accuracy": 0.95,
+      "total_examples": 1400,
+      "n_splits": 22
+    }
+  },
+  "sorl_overall_accuracy": 0.95,
+  "sft_overall_accuracy": 0.5278571428571428
+}

add_sub_sorl_v1_abs10_25K/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0d40fff70ff6d0020577cd516415675f4669c580b180636378d076a80f6caede
+size 650303660

add_sub_sorl_v1_abs10_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": 4,
+  "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_K4_25K",
+  "device": "cuda",
+  "push_to_hub": true,
+  "no_wandb": false,
+  "n_params": 162499262,
+  "run_name": "add_sub_sorl_v1_abs10_25K",
+  "git_commit": "800625019270114adcda289bbd550c4f1109a514",
+  "timestamp": "2026-04-12T02:37:37.086430+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": "iblkjg9n",
+  "wandb_url": "https://wandb.ai/nlp_and_interpretability/sorl-arithmetic/runs/iblkjg9n",
+  "final_accuracy": 0.95,
+  "sft_accuracy": 0.5278571428571428,
+  "eval_method": "ArithmeticEvaluator"
+}