Upload add_sub_sorl_v1_abs10_K1_25K_1L2H256d

add_sub_sorl_v1_abs10_K1_25K_1L2H256d/config.json

@@ -0,0 +1,36 @@
+{
+  "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": 256,
+  "initializer_range": 0.02,
+  "intermediate_size": 1024,
+  "layer_types": [
+    "full_attention"
+  ],
+  "max_position_embeddings": 128,
+  "max_window_layers": 28,
+  "model_type": "qwen3",
+  "num_attention_heads": 2,
+  "num_hidden_layers": 1,
+  "num_key_value_heads": 2,
+  "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_1L2H256d/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_1L2H256d/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": [
+      8.524430274963379,
+      9.103901863098145,
+      11.476999282836914,
+      11.07170295715332,
+      9.617006301879883,
+      8.757355690002441,
+      7.827155113220215,
+      6.2943830490112305,
+      5.568047523498535,
+      4.920413970947266,
+      4.179653167724609,
+      4.015377998352051,
+      3.8744664192199707,
+      3.6563000679016113,
+      3.714816093444824,
+      3.138887882232666,
+      3.476172924041748,
+      3.5434532165527344,
+      3.2397027015686035,
+      3.118135929107666,
+      3.188394546508789,
+      3.03411602973938,
+      3.202545166015625,
+      3.2864937782287598,
+      2.669011116027832,
+      3.0694193840026855,
+      2.755049705505371,
+      2.658294677734375,
+      3.083820343017578,
+      2.523517608642578,
+      2.064281463623047,
+      2.1752371788024902,
+      2.3656363487243652,
+      1.9907090663909912,
+      1.5538188219070435,
+      0.5142228603363037,
+      0.4625096321105957,
+      -0.34851527214050293,
+      -0.9209808111190796,
+      -0.848351240158081,
+      -1.2908415794372559,
+      -1.9059382677078247,
+      -2.5187909603118896,
+      -2.6178600788116455,
+      -2.9667201042175293,
+      -3.4197542667388916,
+      -3.4599928855895996,
+      -4.1904425621032715,
+      -5.044626235961914,
+      -4.426793098449707,
+      -4.773674011230469,
+      -5.006862640380859,
+      -5.084996223449707,
+      -5.224812030792236,
+      -5.75303316116333,
+      -5.572091102600098,
+      -5.412012100219727,
+      -5.735694408416748,
+      -4.803545951843262,
+      -4.7789177894592285,
+      -4.737234115600586,
+      -4.012823104858398,
+      -3.9731242656707764,
+      -4.444588661193848,
+      -4.273018836975098,
+      -4.604288101196289,
+      -4.26378059387207,
+      -4.285333156585693,
+      -3.932020902633667,
+      -4.814791202545166
+    ],
+    "base_loss": [
+      11.692604064941406,
+      10.586286544799805,
+      9.575356483459473,
+      8.626004219055176,
+      7.766415119171143,
+      6.867415904998779,
+      5.944822788238525,
+      4.488890647888184,
+      3.712851047515869,
+      3.1012394428253174,
+      2.6632230281829834,
+      2.547546863555908,
+      2.3669917583465576,
+      2.289210081100464,
+      2.217154026031494,
+      2.0566632747650146,
+      2.032634735107422,
+      1.9611835479736328,
+      2.062990427017212,
+      1.9896939992904663,
+      1.9653056859970093,
+      1.9700582027435303,
+      1.9406718015670776,
+      1.9608590602874756,
+      1.9188401699066162,
+      1.9645260572433472,
+      1.8941099643707275,
+      1.8756904602050781,
+      1.9264477491378784,
+      1.9439666271209717,
+      1.8485826253890991,
+      1.9204241037368774,
+      1.8739402294158936,
+      1.8911492824554443,
+      1.90359628200531,
+      1.8942089080810547,
+      1.9322471618652344,
+      1.8236247301101685,
+      1.8280603885650635,
+      1.8796042203903198,
+      1.8643499612808228,
+      1.9163367748260498,
+      1.8653056621551514,
+      1.8866708278656006,
+      1.8857192993164062,
+      1.8382318019866943,
+      1.8878859281539917,
+      1.8952440023422241,
+      1.9479080438613892,
+      1.8315229415893555,
+      1.853525161743164,
+      1.7897322177886963,
+      1.7672021389007568,
+      1.67568838596344,
+      1.732727289199829,
+      1.6849744319915771,
+      1.6561976671218872,
+      1.6894086599349976,
+      1.5698803663253784,
+      1.526742935180664,
+      1.5828590393066406,
+      1.4620859622955322,
+      1.427542805671692,
+      1.5105113983154297,
+      1.487876296043396,
+      1.4823344945907593,
+      1.458412766456604,
+      1.479615330696106,
+      1.3868151903152466,
+      1.540739893913269
+    ],
+    "info_loss": [
+      -1.2530698776245117,
+      -0.9802732467651367,
+      -0.4201974868774414,
+      -0.18035411834716797,
+      -0.13341093063354492,
+      -0.07659339904785156,
+      -0.04645824432373047,
+      -0.028890609741210938,
+      -0.016974687576293945,
+      -0.016387462615966797,
+      -0.04442238807678223,
+      -0.047545671463012695,
+      -0.042586565017700195,
+      -0.056116580963134766,
+      -0.04165053367614746,
+      -0.08310163021087646,
+      -0.04619729518890381,
+      -0.03194534778594971,
+      -0.07237017154693604,
+      -0.076835036277771,
+      -0.06701540946960449,
+      -0.08263790607452393,
+      -0.06275796890258789,
+      -0.05613744258880615,
+      -0.11362743377685547,
+      -0.07795798778533936,
+      -0.10230779647827148,
+      -0.11010193824768066,
+      -0.07229125499725342,
+      -0.13019275665283203,
+      -0.16647255420684814,
+      -0.16245436668395996,
+      -0.138879656791687,
+      -0.17805254459381104,
+      -0.22295153141021729,
+      -0.32607901096343994,
+      -0.3350032567977905,
+      -0.40525615215301514,
+      -0.4629077911376953,
+      -0.46074652671813965,
+      -0.5033977031707764,
+      -0.5704048871994019,
+      -0.6261879205703735,
+      -0.6387062072753906,
+      -0.6733493804931641,
+      -0.7140244245529175,
+      -0.7230769395828247,
+      -0.796980619430542,
+      -0.887682318687439,
+      -0.8138977289199829,
+      -0.8508479595184326,
+      -0.8680267930030823,
+      -0.8732814192771912,
+      -0.8783826231956482,
+      -0.9365927577018738,
+      -0.9135531187057495,
+      -0.8946322798728943,
+      -0.9301126599311829,
+      -0.8245593905448914,
+      -0.8180517554283142,
+      -0.8192586898803711,
+      -0.734249472618103,
+      -0.7269594073295593,
+      -0.7820543646812439,
+      -0.7622780799865723,
+      -0.7947943210601807,
+      -0.7579862475395203,
+      -0.762677013874054,
+      -0.7179456353187561,
+      -0.821166455745697
+    ],
+    "abs_loss": [
+      2.301368474960327,
+      2.272613763809204,
+      2.1877543926239014,
+      2.0855491161346436,
+      2.003074884414673,
+      1.9530082941055298,
+      1.8817201852798462,
+      1.862039566040039,
+      1.8272932767868042,
+      1.8210479021072388,
+      1.8502413034439087,
+      1.8328882455825806,
+      1.8329845666885376,
+      1.8578821420669556,
+      1.8351255655288696,
+      1.8775984048843384,
+      1.8421878814697266,
+      1.8364969491958618,
+      1.8568716049194336,
+      1.8509222269058228,
+      1.840528130531311,
+      1.8526434898376465,
+      1.8566508293151855,
+      1.8434433937072754,
+      1.8501802682876587,
+      1.8451181650161743,
+      1.8511062860488892,
+      1.8564246892929077,
+      1.8391847610473633,
+      1.8499897718429565,
+      1.8488394021987915,
+      1.8438078165054321,
+      1.8546184301376343,
+      1.857653021812439,
+      1.8427098989486694,
+      1.837051272392273,
+      1.8287359476089478,
+      1.8250846862792969,
+      1.8261111974716187,
+      1.8157788515090942,
+      1.813087821006775,
+      1.8383451700210571,
+      1.7974052429199219,
+      1.8346837759017944,
+      1.8273576498031616,
+      1.8329792022705078,
+      1.8257931470870972,
+      1.8227308988571167,
+      1.8331661224365234,
+      1.8032366037368774,
+      1.8001683950424194,
+      1.8143806457519531,
+      1.8048582077026367,
+      1.8383716344833374,
+      1.8124042749404907,
+      1.7793735265731812,
+      1.7747602462768555,
+      1.7881097793579102,
+      1.7533124685287476,
+      1.776807188987732,
+      1.735535979270935,
+      1.737404465675354,
+      1.7615875005722046,
+      1.7293590307235718,
+      1.7104082107543945,
+      1.7312254905700684,
+      1.700640082359314,
+      1.708137035369873,
+      1.7061353921890259,
+      1.6751855611801147
+    ],
+    "zipf_loss": [
+      9.1323881149292,
+      8.093086242675781,
+      5.884842872619629,
+      4.040685653686523,
+      2.9843926429748535,
+      2.4605727195739746,
+      2.158742666244507,
+      1.9081945419311523,
+      1.8422138690948486,
+      1.8009443283081055,
+      1.775629997253418,
+      1.75999915599823,
+      1.7500418424606323,
+      1.7424676418304443,
+      1.7306548357009888,
+      1.7254811525344849,
+      1.7212923765182495,
+      1.7180733680725098,
+      1.7147266864776611,
+      1.711700201034546,
+      1.709190011024475,
+      1.7051725387573242,
+      1.703787922859192,
+      1.702664852142334,
+      1.7014273405075073,
+      1.699961543083191,
+      1.6989071369171143,
+      1.6979811191558838,
+      1.696366548538208,
+      1.6964795589447021,
+      1.695540428161621,
+      1.6949759721755981,
+      1.6950308084487915,
+      1.6943199634552002,
+      1.6954668760299683,
+      1.697098970413208,
+      1.6974214315414429,
+      1.697913408279419,
+      1.6974254846572876,
+      1.6979317665100098,
+      1.697476863861084,
+      1.6979395151138306,
+      1.6980416774749756,
+      1.6990625858306885,
+      1.6983184814453125,
+      1.6989600658416748,
+      1.7003107070922852,
+      1.701846718788147,
+      1.7009726762771606,
+      1.700337290763855,
+      1.7012641429901123,
+      1.7022359371185303,
+      1.7001292705535889,
+      1.6994887590408325,
+      1.6989268064498901,
+      1.7005280256271362,
+      1.7006367444992065,
+      1.6972121000289917,
+      1.6968367099761963,
+      1.6971760988235474,
+      1.6989400386810303,
+      1.6938459873199463,
+      1.6927683353424072,
+      1.6925078630447388,
+      1.6908442974090576,
+      1.688198208808899,
+      1.687605381011963,
+      1.6910079717636108,
+      1.6900064945220947,
+      1.6886142492294312
+    ],
+    "denoise_loss": [],
+    "ortho_loss": [
+      0.7410624623298645,
+      0.5486676096916199,
+      0.4325517416000366,
+      0.4293130338191986,
+      0.42301344871520996,
+      0.43677014112472534,
+      0.4322786331176758,
+      0.43111148476600647,
+      0.42917412519454956,
+      0.4202711284160614,
+      0.40449219942092896,
+      0.38009271025657654,
+      0.38311055302619934,
+      0.38355082273483276,
+      0.3772588074207306,
+      0.37517693638801575,
+      0.35672980546951294,
+      0.3655826151371002,
+      0.36484745144844055,
+      0.36458733677864075,
+      0.3716451823711395,
+      0.3796772062778473,
+      0.37268635630607605,
+      0.36658456921577454,
+      0.36402279138565063,
+      0.36522912979125977,
+      0.3712886571884155,
+      0.3701127767562866,
+      0.3669368028640747,
+      0.35840272903442383,
+      0.36034736037254333,
+      0.36091148853302,
+      0.35899659991264343,
+      0.356486439704895,
+      0.3690042495727539,
+      0.3797115981578827,
+      0.3873102366924286,
+      0.39992478489875793,
+      0.41160136461257935,
+      0.42042696475982666,
+      0.42867836356163025,
+      0.43672993779182434,
+      0.4450317621231079,
+      0.45142629742622375,
+      0.451881468296051,
+      0.45120909810066223,
+      0.4502365291118622,
+      0.44785618782043457,
+      0.4534400999546051,
+      0.45658400654792786,
+      0.45758968591690063,
+      0.45804986357688904,
+      0.4550400674343109,
+      0.451234370470047,
+      0.4503278136253357,
+      0.45379236340522766,
+      0.4566922187805176,
+      0.4583297073841095,
+      0.46125417947769165,
+      0.46176251769065857,
+      0.4625169634819031,
+      0.46337494254112244,
+      0.46481263637542725,
+      0.4634028375148773,
+      0.4632355272769928,
+      0.46351659297943115,
+      0.46328744292259216,
+      0.4644387364387512,
+      0.46469375491142273,
+      0.4648001492023468
+    ],
+    "lr": [
+      8.376068376068378e-06,
+      1.6923076923076924e-05,
+      2e-05,
+      2e-05,
+      2e-05,
+      2e-05,
+      2e-05,
+      2e-05,
+      2e-05,
+      2e-05,
+      2e-05,
+      2e-05,
+      2e-05,
+      2e-05,
+      2e-05,
+      2e-05,
+      2e-05,
+      2e-05,
+      2e-05,
+      2e-05,
+      2e-05,
+      2e-05,
+      2e-05,
+      2e-05,
+      2e-05,
+      2e-05,
+      2e-05,
+      2e-05,
+      2e-05,
+      2e-05,
+      2e-05,
+      2e-05,
+      2e-05,
+      2e-05,
+      2e-05,
+      2e-05,
+      2e-05,
+      2e-05,
+      2e-05,
+      2e-05,
+      2e-05,
+      2e-05,
+      1.9973899288162407e-05,
+      1.9380701291853413e-05,
+      1.8787503295544426e-05,
+      1.8194305299235432e-05,
+      1.7601107302926442e-05,
+      1.7007909306617455e-05,
+      1.6414711310308464e-05,
+      1.5335090957026102e-05,
+      1.474189296071711e-05,
+      1.4148694964408121e-05,
+      1.355549696809913e-05,
+      1.2962298971790142e-05,
+      1.2369100975481152e-05,
+      1.1775902979172158e-05,
+      1.0696282625889797e-05,
+      1.0103084629580805e-05,
+      9.50988663327182e-06,
+      8.91668863696283e-06,
+      8.323490640653837e-06,
+      7.730292644344845e-06,
+      7.137094648035855e-06,
+      6.057474294753492e-06,
+      5.4642762984445065e-06,
+      4.871078302135514e-06,
+      4.277880305826523e-06,
+      3.684682309517532e-06,
+      3.091484313208541e-06,
+      2.4982863168995496e-06
+    ],
+    "emb_lr": [],
+    "eval_step": [
+      350,
+      741,
+      1132,
+      1523,
+      1914,
+      2305,
+      2696,
+      3087,
+      3478,
+      3869
+    ],
+    "eval_accuracy": [
+      0.01,
+      0.0,
+      0.0,
+      0.01,
+      0.0,
+      0.0,
+      0.01,
+      0.0,
+      0.0,
+      0.0
+    ]
+  },
+  "final_accuracy": 0.06291666666666666,
+  "sft_eval": {
+    "config": {
+      "ops": "add_sub",
+      "K": null,
+      "mode": "sft",
+      "n_digits": 6,
+      "n_per_split": 100
+    },
+    "splits": {
+      "add_S0": {
+        "full_accuracy": 0.01,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 0.5652892561983471,
+            "count": 605
+          },
+          "SS": {
+            "accuracy": 0.7578947368421053,
+            "count": 95
+          }
+        }
+      },
+      "add_S1": {
+        "full_accuracy": 0.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 0.5735294117647058,
+            "count": 204
+          },
+          "SC": {
+            "accuracy": 0.41420118343195267,
+            "count": 169
+          },
+          "SS": {
+            "accuracy": 0.5483870967741935,
+            "count": 31
+          },
+          "UC": {
+            "accuracy": 0.38175675675675674,
+            "count": 296
+          }
+        }
+      },
+      "add_S2": {
+        "full_accuracy": 0.02,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 0.5828220858895705,
+            "count": 163
+          },
+          "SC": {
+            "accuracy": 0.38461538461538464,
+            "count": 130
+          },
+          "SS": {
+            "accuracy": 0.7586206896551724,
+            "count": 87
+          },
+          "UC": {
+            "accuracy": 0.4433497536945813,
+            "count": 203
+          },
+          "US": {
+            "accuracy": 0.46153846153846156,
+            "count": 117
+          }
+        }
+      },
+      "add_S3": {
+        "full_accuracy": 0.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 0.6694214876033058,
+            "count": 121
+          },
+          "SC": {
+            "accuracy": 0.371900826446281,
+            "count": 121
+          },
+          "SS": {
+            "accuracy": 0.8775510204081632,
+            "count": 49
+          },
+          "UC": {
+            "accuracy": 0.3333333333333333,
+            "count": 186
+          },
+          "US": {
+            "accuracy": 0.23766816143497757,
+            "count": 223
+          }
+        }
+      },
+      "add_S4": {
+        "full_accuracy": 0.05,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 0.7019230769230769,
+            "count": 104
+          },
+          "SC": {
+            "accuracy": 0.3584905660377358,
+            "count": 106
+          },
+          "SS": {
+            "accuracy": 0.8695652173913043,
+            "count": 23
+          },
+          "UC": {
+            "accuracy": 0.39375,
+            "count": 160
+          },
+          "US": {
+            "accuracy": 0.24104234527687296,
+            "count": 307
+          }
+        }
+      },
+      "add_S5": {
+        "full_accuracy": 0.04,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 0.65,
+            "count": 100
+          },
+          "SC": {
+            "accuracy": 0.28,
+            "count": 100
+          },
+          "UC": {
+            "accuracy": 0.29,
+            "count": 100
+          },
+          "US": {
+            "accuracy": 0.22,
+            "count": 400
+          }
+        }
+      },
+      "add_S6": {
+        "full_accuracy": 0.09,
+        "n_examples": 100,
+        "per_subtask": {
+          "SC": {
+            "accuracy": 0.34,
+            "count": 100
+          },
+          "UC": {
+            "accuracy": 0.43,
+            "count": 100
+          },
+          "US": {
+            "accuracy": 0.32,
+            "count": 500
+          }
+        }
+      },
+      "add_random": {
+        "full_accuracy": 0.005,
+        "n_examples": 200,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 0.5682326621923938,
+            "count": 447
+          },
+          "SC": {
+            "accuracy": 0.40625,
+            "count": 320
+          },
+          "SS": {
+            "accuracy": 0.8035714285714286,
+            "count": 56
+          },
+          "UC": {
+            "accuracy": 0.34026465028355385,
+            "count": 529
+          },
+          "US": {
+            "accuracy": 0.2916666666666667,
+            "count": 48
+          }
+        }
+      },
+      "add_C3": {
+        "full_accuracy": 0.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 0.63,
+            "count": 300
+          },
+          "SC": {
+            "accuracy": 0.34,
+            "count": 100
+          },
+          "UC": {
+            "accuracy": 0.29533678756476683,
+            "count": 193
+          },
+          "US": {
+            "accuracy": 0.29906542056074764,
+            "count": 107
+          }
+        }
+      },
+      "add_C4": {
+        "full_accuracy": 0.01,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 0.695,
+            "count": 200
+          },
+          "SC": {
+            "accuracy": 0.4,
+            "count": 100
+          },
+          "UC": {
+            "accuracy": 0.26171875,
+            "count": 256
+          },
+          "US": {
+            "accuracy": 0.2777777777777778,
+            "count": 144
+          }
+        }
+      },
+      "add_C5": {
+        "full_accuracy": 0.03,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 0.74,
+            "count": 100
+          },
+          "SC": {
+            "accuracy": 0.49,
+            "count": 100
+          },
+          "UC": {
+            "accuracy": 0.30718954248366015,
+            "count": 306
+          },
+          "US": {
+            "accuracy": 0.4742268041237113,
+            "count": 194
+          }
+        }
+      },
+      "add_C6": {
+        "full_accuracy": 0.01,
+        "n_examples": 100,
+        "per_subtask": {
+          "SC": {
+            "accuracy": 0.46,
+            "count": 100
+          },
+          "UC": {
+            "accuracy": 0.37158469945355194,
+            "count": 366
+          },
+          "US": {
+            "accuracy": 0.47435897435897434,
+            "count": 234
+          }
+        }
+      },
+      "sub_M0": {
+        "full_accuracy": 0.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 0.3876871880199667,
+            "count": 601
+          },
+          "ME": {
+            "accuracy": 0.7878787878787878,
+            "count": 99
+          }
+        }
+      },
+      "sub_M1": {
+        "full_accuracy": 0.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 0.5376344086021505,
+            "count": 279
+          },
+          "MB": {
+            "accuracy": 0.11724137931034483,
+            "count": 145
+          },
+          "ME": {
+            "accuracy": 0.7916666666666666,
+            "count": 24
+          },
+          "UB": {
+            "accuracy": 0.14682539682539683,
+            "count": 252
+          }
+        }
+      },
+      "sub_M2": {
+        "full_accuracy": 0.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 0.7417840375586855,
+            "count": 213
+          },
+          "MB": {
+            "accuracy": 0.061946902654867256,
+            "count": 113
+          },
+          "ME": {
+            "accuracy": 0.8235294117647058,
+            "count": 85
+          },
+          "UB": {
+            "accuracy": 0.143646408839779,
+            "count": 181
+          },
+          "UD": {
+            "accuracy": 0.07407407407407407,
+            "count": 108
+          }
+        }
+      },
+      "sub_M3": {
+        "full_accuracy": 0.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 0.8044692737430168,
+            "count": 179
+          },
+          "MB": {
+            "accuracy": 0.02912621359223301,
+            "count": 103
+          },
+          "ME": {
+            "accuracy": 0.9107142857142857,
+            "count": 56
+          },
+          "UB": {
+            "accuracy": 0.2214765100671141,
+            "count": 149
+          },
+          "UD": {
+            "accuracy": 0.1267605633802817,
+            "count": 213
+          }
+        }
+      },
+      "sub_M4": {
+        "full_accuracy": 0.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 0.615,
+            "count": 200
+          },
+          "MB": {
+            "accuracy": 0.11,
+            "count": 100
+          },
+          "UB": {
+            "accuracy": 0.36,
+            "count": 100
+          },
+          "UD": {
+            "accuracy": 0.14,
+            "count": 300
+          }
+        }
+      },
+      "sub_M5": {
+        "full_accuracy": 0.01,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "MB": {
+            "accuracy": 0.08,
+            "count": 100
+          },
+          "UB": {
+            "accuracy": 0.43,
+            "count": 100
+          },
+          "UD": {
+            "accuracy": 0.045,
+            "count": 400
+          }
+        }
+      },
+      "sub_random": {
+        "full_accuracy": 0.0,
+        "n_examples": 200,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 0.54,
+            "count": 600
+          },
+          "MB": {
+            "accuracy": 0.0898876404494382,
+            "count": 267
+          },
+          "ME": {
+            "accuracy": 0.8301886792452831,
+            "count": 53
+          },
+          "UB": {
+            "accuracy": 0.14578587699316628,
+            "count": 439
+          },
+          "UD": {
+            "accuracy": 0.0975609756097561,
+            "count": 41
+          }
+        }
+      },
+      "sub_B3": {
+        "full_accuracy": 0.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 0.44666666666666666,
+            "count": 300
+          },
+          "MB": {
+            "accuracy": 0.01,
+            "count": 100
+          },
+          "UB": {
+            "accuracy": 0.18781725888324874,
+            "count": 197
+          },
+          "UD": {
+            "accuracy": 0.0970873786407767,
+            "count": 103
+          }
+        }
+      },
+      "sub_B4": {
+        "full_accuracy": 0.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 0.65,
+            "count": 200
+          },
+          "MB": {
+            "accuracy": 0.02,
+            "count": 100
+          },
+          "UB": {
+            "accuracy": 0.1659919028340081,
+            "count": 247
+          },
+          "UD": {
+            "accuracy": 0.08496732026143791,
+            "count": 153
+          }
+        }
+      },
+      "sub_B5": {
+        "full_accuracy": 0.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "MB": {
+            "accuracy": 0.04,
+            "count": 100
+          },
+          "UB": {
+            "accuracy": 0.18791946308724833,
+            "count": 298
+          },
+          "UD": {
+            "accuracy": 0.06930693069306931,
+            "count": 202
+          }
+        }
+      }
+    },
+    "summary": {
+      "overall_accuracy": 0.011666666666666667,
+      "total_examples": 2400,
+      "n_splits": 22
+    }
+  },
+  "sorl_eval": {
+    "config": {
+      "ops": "add_sub",
+      "K": 1,
+      "mode": "sorl",
+      "n_digits": 6,
+      "n_per_split": 100
+    },
+    "splits": {
+      "add_S0": {
+        "full_accuracy": 0.65,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 0.9388429752066115,
+            "count": 605
+          },
+          "SS": {
+            "accuracy": 0.8736842105263158,
+            "count": 95
+          }
+        }
+      },
+      "add_S1": {
+        "full_accuracy": 0.01,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 0.9509803921568627,
+            "count": 204
+          },
+          "SC": {
+            "accuracy": 0.8520710059171598,
+            "count": 169
+          },
+          "SS": {
+            "accuracy": 0.9032258064516129,
+            "count": 31
+          },
+          "UC": {
+            "accuracy": 0.24662162162162163,
+            "count": 296
+          }
+        }
+      },
+      "add_S2": {
+        "full_accuracy": 0.02,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 0.9079754601226994,
+            "count": 163
+          },
+          "SC": {
+            "accuracy": 0.8538461538461538,
+            "count": 130
+          },
+          "SS": {
+            "accuracy": 0.896551724137931,
+            "count": 87
+          },
+          "UC": {
+            "accuracy": 0.22167487684729065,
+            "count": 203
+          },
+          "US": {
+            "accuracy": 0.21367521367521367,
+            "count": 117
+          }
+        }
+      },
+      "add_S3": {
+        "full_accuracy": 0.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 0.9256198347107438,
+            "count": 121
+          },
+          "SC": {
+            "accuracy": 0.8264462809917356,
+            "count": 121
+          },
+          "SS": {
+            "accuracy": 0.8979591836734694,
+            "count": 49
+          },
+          "UC": {
+            "accuracy": 0.3172043010752688,
+            "count": 186
+          },
+          "US": {
+            "accuracy": 0.08520179372197309,
+            "count": 223
+          }
+        }
+      },
+      "add_S4": {
+        "full_accuracy": 0.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 0.9711538461538461,
+            "count": 104
+          },
+          "SC": {
+            "accuracy": 0.8113207547169812,
+            "count": 106
+          },
+          "SS": {
+            "accuracy": 0.9565217391304348,
+            "count": 23
+          },
+          "UC": {
+            "accuracy": 0.3375,
+            "count": 160
+          },
+          "US": {
+            "accuracy": 0.10749185667752444,
+            "count": 307
+          }
+        }
+      },
+      "add_S5": {
+        "full_accuracy": 0.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 0.94,
+            "count": 100
+          },
+          "SC": {
+            "accuracy": 0.63,
+            "count": 100
+          },
+          "UC": {
+            "accuracy": 0.05,
+            "count": 100
+          },
+          "US": {
+            "accuracy": 0.0825,
+            "count": 400
+          }
+        }
+      },
+      "add_S6": {
+        "full_accuracy": 0.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "SC": {
+            "accuracy": 0.54,
+            "count": 100
+          },
+          "UC": {
+            "accuracy": 0.21,
+            "count": 100
+          },
+          "US": {
+            "accuracy": 0.046,
+            "count": 500
+          }
+        }
+      },
+      "add_random": {
+        "full_accuracy": 0.035,
+        "n_examples": 200,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 0.9194630872483222,
+            "count": 447
+          },
+          "SC": {
+            "accuracy": 0.8625,
+            "count": 320
+          },
+          "SS": {
+            "accuracy": 0.8571428571428571,
+            "count": 56
+          },
+          "UC": {
+            "accuracy": 0.2608695652173913,
+            "count": 529
+          },
+          "US": {
+            "accuracy": 0.1875,
+            "count": 48
+          }
+        }
+      },
+      "add_C3": {
+        "full_accuracy": 0.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 0.95,
+            "count": 300
+          },
+          "SC": {
+            "accuracy": 0.81,
+            "count": 100
+          },
+          "UC": {
+            "accuracy": 0.20207253886010362,
+            "count": 193
+          },
+          "US": {
+            "accuracy": 0.14953271028037382,
+            "count": 107
+          }
+        }
+      },
+      "add_C4": {
+        "full_accuracy": 0.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 0.935,
+            "count": 200
+          },
+          "SC": {
+            "accuracy": 0.76,
+            "count": 100
+          },
+          "UC": {
+            "accuracy": 0.19140625,
+            "count": 256
+          },
+          "US": {
+            "accuracy": 0.19444444444444445,
+            "count": 144
+          }
+        }
+      },
+      "add_C5": {
+        "full_accuracy": 0.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 0.94,
+            "count": 100
+          },
+          "SC": {
+            "accuracy": 0.81,
+            "count": 100
+          },
+          "UC": {
+            "accuracy": 0.19607843137254902,
+            "count": 306
+          },
+          "US": {
+            "accuracy": 0.10309278350515463,
+            "count": 194
+          }
+        }
+      },
+      "add_C6": {
+        "full_accuracy": 0.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "SC": {
+            "accuracy": 0.62,
+            "count": 100
+          },
+          "UC": {
+            "accuracy": 0.31693989071038253,
+            "count": 366
+          },
+          "US": {
+            "accuracy": 0.14957264957264957,
+            "count": 234
+          }
+        }
+      },
+      "sub_M0": {
+        "full_accuracy": 0.66,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 0.9384359400998337,
+            "count": 601
+          },
+          "ME": {
+            "accuracy": 0.9696969696969697,
+            "count": 99
+          }
+        }
+      },
+      "sub_M1": {
+        "full_accuracy": 0.01,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 0.9283154121863799,
+            "count": 279
+          },
+          "MB": {
+            "accuracy": 0.8,
+            "count": 145
+          },
+          "ME": {
+            "accuracy": 1.0,
+            "count": 24
+          },
+          "UB": {
+            "accuracy": 0.12301587301587301,
+            "count": 252
+          }
+        }
+      },
+      "sub_M2": {
+        "full_accuracy": 0.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 0.9765258215962441,
+            "count": 213
+          },
+          "MB": {
+            "accuracy": 0.7699115044247787,
+            "count": 113
+          },
+          "ME": {
+            "accuracy": 0.9882352941176471,
+            "count": 85
+          },
+          "UB": {
+            "accuracy": 0.15469613259668508,
+            "count": 181
+          },
+          "UD": {
+            "accuracy": 0.027777777777777776,
+            "count": 108
+          }
+        }
+      },
+      "sub_M3": {
+        "full_accuracy": 0.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 0.9832402234636871,
+            "count": 179
+          },
+          "MB": {
+            "accuracy": 0.6796116504854369,
+            "count": 103
+          },
+          "ME": {
+            "accuracy": 0.9642857142857143,
+            "count": 56
+          },
+          "UB": {
+            "accuracy": 0.16778523489932887,
+            "count": 149
+          },
+          "UD": {
+            "accuracy": 0.06572769953051644,
+            "count": 213
+          }
+        }
+      },
+      "sub_M4": {
+        "full_accuracy": 0.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 0.95,
+            "count": 200
+          },
+          "MB": {
+            "accuracy": 0.68,
+            "count": 100
+          },
+          "UB": {
+            "accuracy": 0.16,
+            "count": 100
+          },
+          "UD": {
+            "accuracy": 0.05333333333333334,
+            "count": 300
+          }
+        }
+      },
+      "sub_M5": {
+        "full_accuracy": 0.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "MB": {
+            "accuracy": 0.45,
+            "count": 100
+          },
+          "UB": {
+            "accuracy": 0.2,
+            "count": 100
+          },
+          "UD": {
+            "accuracy": 0.045,
+            "count": 400
+          }
+        }
+      },
+      "sub_random": {
+        "full_accuracy": 0.045,
+        "n_examples": 200,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 0.9166666666666666,
+            "count": 600
+          },
+          "MB": {
+            "accuracy": 0.8389513108614233,
+            "count": 267
+          },
+          "ME": {
+            "accuracy": 0.9811320754716981,
+            "count": 53
+          },
+          "UB": {
+            "accuracy": 0.10022779043280182,
+            "count": 439
+          },
+          "UD": {
+            "accuracy": 0.0,
+            "count": 41
+          }
+        }
+      },
+      "sub_B3": {
+        "full_accuracy": 0.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 0.9333333333333333,
+            "count": 300
+          },
+          "MB": {
+            "accuracy": 0.84,
+            "count": 100
+          },
+          "UB": {
+            "accuracy": 0.15736040609137056,
+            "count": 197
+          },
+          "UD": {
+            "accuracy": 0.038834951456310676,
+            "count": 103
+          }
+        }
+      },
+      "sub_B4": {
+        "full_accuracy": 0.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 0.96,
+            "count": 200
+          },
+          "MB": {
+            "accuracy": 0.84,
+            "count": 100
+          },
+          "UB": {
+            "accuracy": 0.13360323886639677,
+            "count": 247
+          },
+          "UD": {
+            "accuracy": 0.06535947712418301,
+            "count": 153
+          }
+        }
+      },
+      "sub_B5": {
+        "full_accuracy": 0.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "MB": {
+            "accuracy": 0.64,
+            "count": 100
+          },
+          "UB": {
+            "accuracy": 0.20469798657718122,
+            "count": 298
+          },
+          "UD": {
+            "accuracy": 0.06930693069306931,
+            "count": 202
+          }
+        }
+      }
+    },
+    "summary": {
+      "overall_accuracy": 0.06291666666666666,
+      "total_examples": 2400,
+      "n_splits": 22
+    }
+  },
+  "sorl_overall_accuracy": 0.06291666666666666,
+  "sft_overall_accuracy": 0.011666666666666667
+}

add_sub_sorl_v1_abs10_K1_25K_1L2H256d/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a10bb1f5b2eb181b8afa8885bc9617b44e94d4a7398d421530a00ee21fe002b3
+size 315091124

add_sub_sorl_v1_abs10_K1_25K_1L2H256d/train_config.json

@@ -0,0 +1,84 @@
+{
+  "num_rollouts": 4,
+  "K": 1,
+  "max_iterations": 2,
+  "memory_span_abs": 1792,
+  "memory_span_traj": 1792,
+  "temperature": 1.0,
+  "ar_search": false,
+  "response_only_abs": false,
+  "alpha_info_gain": 10.0,
+  "alpha_abs": 0.1,
+  "alpha_soft_zipf": 1.0,
+  "alpha_ortho": 0.0,
+  "alpha_anchor": 0.0,
+  "alpha_jacobi": 0.0,
+  "decay": 0.8,
+  "target_vocab_util": 0.8,
+  "min_abs_ppl": 0.0,
+  "zipf_alpha": 1.0,
+  "lr": 2e-05,
+  "emb_lr_mult": 1.0,
+  "weight_decay": 0.01,
+  "warmup_steps": 117,
+  "cooldown_frac": 0.4,
+  "max_grad_norm": 1.0,
+  "vq_abs_pretrain_steps": 0,
+  "vq_abs_pretrain_lr": 0.001,
+  "vq_abs_pretrain_layer": -1,
+  "vq_abs_pretrain_batch_size": 256,
+  "vq_abs_pretrain_target_vectors": 20000,
+  "batch_size": 64,
+  "gradient_accumulation_steps": 1,
+  "num_epochs": 10,
+  "emb_warmup_steps": 0,
+  "log_every": 50,
+  "eval_every": 390,
+  "save_every": 999999,
+  "eval_samples": 100,
+  "output_dir": "ckpt/sweep/as_sorl_abs10_K1_25K_1L2H256d",
+  "eval_K": 4,
+  "alpha_traj": 0.0,
+  "corrupt_method": "shuffle",
+  "corrupt_ratio": 0.3,
+  "alpha_contrastive": 1.0,
+  "gamma_contrastive": 0.5,
+  "alpha_masked_traj": 0.0,
+  "mask_nl_ratio": 0.3,
+  "mask_nl_mode": "fixed",
+  "mask_nl_fixed_id": 0,
+  "use_ste": true,
+  "n_inner": 1,
+  "random_K": null,
+  "strip_suffix": null,
+  "compress_prefix": null,
+  "random_mem_span": null,
+  "warmup_ratio": 0.03,
+  "beta2": 0.999,
+  "seed": 42,
+  "n_digits": 6,
+  "n_layer": 1,
+  "n_head": 2,
+  "n_embd": 256,
+  "ops": "add_sub",
+  "abs_vocab": 10,
+  "dataset_size": 25000,
+  "mode": "sorl",
+  "device": "cuda",
+  "push_to_hub": true,
+  "no_wandb": false,
+  "n_params": 78696448,
+  "run_name": "add_sub_sorl_v1_abs10_K1_25K_1L2H256d",
+  "git_commit": "3ae8ca0d6b88706715f25991b1e1acd0e3a6e0a6",
+  "timestamp": "2026-04-12T20:10:10.507089+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": "8olk3u7a",
+  "wandb_url": "https://wandb.ai/nlp_and_interpretability/sorl-arithmetic/runs/8olk3u7a",
+  "final_accuracy": 0.06291666666666666,
+  "sft_accuracy": 0.011666666666666667,
+  "eval_method": "ArithmeticEvaluator"
+}