Upload add_sub_sorl_v1_abs10_K1_25K_1L3H510d

add_sub_sorl_v1_abs10_K1_25K_1L3H510d/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": 510,
+  "initializer_range": 0.02,
+  "intermediate_size": 2040,
+  "layer_types": [
+    "full_attention"
+  ],
+  "max_position_embeddings": 128,
+  "max_window_layers": 28,
+  "model_type": "qwen3",
+  "num_attention_heads": 3,
+  "num_hidden_layers": 1,
+  "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_1L3H510d/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_1L3H510d/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": [
+      -2.59329891204834,
+      7.806329250335693,
+      4.891926288604736,
+      3.912823438644409,
+      3.36497163772583,
+      3.413069248199463,
+      3.1604514122009277,
+      3.0855469703674316,
+      2.895873546600342,
+      2.8154397010803223,
+      2.8238906860351562,
+      2.5073657035827637,
+      1.2554247379302979,
+      -0.46714138984680176,
+      -5.484283447265625,
+      -6.567154884338379,
+      -6.758439064025879,
+      -7.20958137512207,
+      -8.60386848449707,
+      -9.11019515991211,
+      -9.290766716003418,
+      -9.791807174682617,
+      -10.01004695892334,
+      -10.105368614196777,
+      -9.49570083618164,
+      -10.431818008422852,
+      -10.326833724975586,
+      -10.126893997192383,
+      -10.160858154296875,
+      -10.941141128540039,
+      -10.631235122680664,
+      -10.644250869750977,
+      -10.591297149658203,
+      -11.348092079162598,
+      -11.023357391357422,
+      -11.110491752624512,
+      -11.359291076660156,
+      -11.153099060058594,
+      -10.987231254577637,
+      -11.629621505737305,
+      -10.848411560058594,
+      -11.752678871154785,
+      -11.179244995117188,
+      -11.764345169067383,
+      -11.807055473327637,
+      -11.234673500061035,
+      -12.306229591369629,
+      -11.79053783416748,
+      -12.127859115600586,
+      -11.52390193939209,
+      -11.549543380737305,
+      -13.031827926635742,
+      -12.10395336151123,
+      -11.681001663208008,
+      -12.33035945892334,
+      -12.12756633758545,
+      -12.05741024017334,
+      -12.84146785736084,
+      -12.221896171569824,
+      -12.554245948791504,
+      -12.93739128112793,
+      -12.630762100219727,
+      -12.096063613891602,
+      -12.651780128479004,
+      -13.059272766113281,
+      -12.436356544494629,
+      -12.79887580871582,
+      -12.472254753112793,
+      -11.984771728515625,
+      -13.45839786529541
+    ],
+    "base_loss": [
+      10.196748733520508,
+      6.539699077606201,
+      4.092434406280518,
+      2.346903085708618,
+      2.0746912956237793,
+      1.9594690799713135,
+      1.8837043046951294,
+      1.8864774703979492,
+      1.8490782976150513,
+      1.7932301759719849,
+      1.7640596628189087,
+      1.8651959896087646,
+      1.8547264337539673,
+      1.886150598526001,
+      1.9799963235855103,
+      1.80038583278656,
+      1.8168869018554688,
+      1.7473572492599487,
+      1.8770685195922852,
+      1.8492628335952759,
+      1.8485273122787476,
+      1.834829330444336,
+      1.843698263168335,
+      1.8508756160736084,
+      1.804511308670044,
+      1.85122549533844,
+      1.8213146924972534,
+      1.8022531270980835,
+      1.8216105699539185,
+      1.8896723985671997,
+      1.7682313919067383,
+      1.851951003074646,
+      1.8244448900222778,
+      1.860870122909546,
+      1.844719648361206,
+      1.8215335607528687,
+      1.8681211471557617,
+      1.7685410976409912,
+      1.77374267578125,
+      1.83017098903656,
+      1.780069351196289,
+      1.8695632219314575,
+      1.803731918334961,
+      1.794657588005066,
+      1.8132097721099854,
+      1.744333028793335,
+      1.850972056388855,
+      1.79466712474823,
+      1.865286111831665,
+      1.775697112083435,
+      1.831557273864746,
+      1.8576253652572632,
+      1.7820489406585693,
+      1.7218695878982544,
+      1.7814966440200806,
+      1.771523356437683,
+      1.748187780380249,
+      1.839253544807434,
+      1.7425659894943237,
+      1.7759649753570557,
+      1.8560504913330078,
+      1.794585108757019,
+      1.7335294485092163,
+      1.806968092918396,
+      1.8292748928070068,
+      1.7885210514068604,
+      1.8298585414886475,
+      1.7961997985839844,
+      1.723584532737732,
+      1.8568757772445679
+    ],
+    "info_loss": [
+      -2.1446304321289062,
+      -0.32575225830078125,
+      -0.1383652687072754,
+      -0.03788328170776367,
+      -0.06099653244018555,
+      -0.043364644050598145,
+      -0.060524702072143555,
+      -0.06769096851348877,
+      -0.08283662796020508,
+      -0.08506667613983154,
+      -0.08107435703277588,
+      -0.12192976474761963,
+      -0.23974978923797607,
+      -0.40476346015930176,
+      -0.9116840362548828,
+      -1.002124309539795,
+      -1.0228478908538818,
+      -1.060563087463379,
+      -1.2133022546768188,
+      -1.2610199451446533,
+      -1.2785308361053467,
+      -1.326535940170288,
+      -1.3484688997268677,
+      -1.3588534593582153,
+      -1.2928407192230225,
+      -1.3912382125854492,
+      -1.3773154020309448,
+      -1.3551386594772339,
+      -1.3596680164337158,
+      -1.44418203830719,
+      -1.3997809886932373,
+      -1.4097068309783936,
+      -1.400420069694519,
+      -1.4767446517944336,
+      -1.4412634372711182,
+      -1.4422601461410522,
+      -1.468010663986206,
+      -1.437137484550476,
+      -1.421332597732544,
+      -1.4890285730361938,
+      -1.4050776958465576,
+      -1.5023729801177979,
+      -1.4319555759429932,
+      -1.4917011260986328,
+      -1.4962141513824463,
+      -1.4313055276870728,
+      -1.5444698333740234,
+      -1.487035870552063,
+      -1.5295355319976807,
+      -1.4528287649154663,
+      -1.4570941925048828,
+      -1.6042754650115967,
+      -1.5032556056976318,
+      -1.4498255252838135,
+      -1.5202281475067139,
+      -1.4980217218399048,
+      -1.488149881362915,
+      -1.574017882347107,
+      -1.498759150505066,
+      -1.5380897521972656,
+      -1.5861152410507202,
+      -1.5452888011932373,
+      -1.4880478382110596,
+      -1.5506173372268677,
+      -1.5903542041778564,
+      -1.5227900743484497,
+      -1.5617271661758423,
+      -1.5277729034423828,
+      -1.4715402126312256,
+      -1.62960946559906
+    ],
+    "abs_loss": [
+      2.2833845615386963,
+      2.1349236965179443,
+      1.895667552947998,
+      1.8790363073349,
+      1.8763422966003418,
+      1.863279938697815,
+      1.8707689046859741,
+      1.8497987985610962,
+      1.8467472791671753,
+      1.8501476049423218,
+      1.8510236740112305,
+      1.7979803085327148,
+      1.6018939018249512,
+      1.4351223707199097,
+      1.3627341985702515,
+      1.3827515840530396,
+      1.378796935081482,
+      1.3818906545639038,
+      1.3893027305603027,
+      1.376089096069336,
+      1.3647276163101196,
+      1.341958999633789,
+      1.3429460525512695,
+      1.2983181476593018,
+      1.3058513402938843,
+      1.3161550760269165,
+      1.269581913948059,
+      1.2819956541061401,
+      1.225730061531067,
+      1.2145700454711914,
+      1.2128896713256836,
+      1.181650161743164,
+      1.1704578399658203,
+      1.100685477256775,
+      1.1276469230651855,
+      1.0256637334823608,
+      0.9807112812995911,
+      0.9855987429618835,
+      1.0340598821640015,
+      0.962256133556366,
+      0.9177753925323486,
+      0.923234760761261,
+      0.9172429442405701,
+      0.9422297477722168,
+      0.9557555317878723,
+      0.959925651550293,
+      0.8680426478385925,
+      0.8737877011299133,
+      0.9473252296447754,
+      0.8711683750152588,
+      0.8707416653633118,
+      0.7836803793907166,
+      0.847979724407196,
+      0.7654786109924316,
+      0.793886661529541,
+      0.7679941058158875,
+      0.7534409165382385,
+      0.7690866589546204,
+      0.7134169936180115,
+      0.7436110973358154,
+      0.743655264377594,
+      0.6757664084434509,
+      0.6782504916191101,
+      0.684099018573761,
+      0.6790602207183838,
+      0.6778237223625183,
+      0.6263924241065979,
+      0.6698463559150696,
+      0.650520384311676,
+      0.6476595401763916
+    ],
+    "zipf_loss": [
+      8.427918434143066,
+      4.310660362243652,
+      1.9935778379440308,
+      1.7568495273590088,
+      1.7126115560531616,
+      1.7009185552597046,
+      1.6949173212051392,
+      1.6909992694854736,
+      1.6904866695404053,
+      1.687861680984497,
+      1.6854722499847412,
+      1.6816692352294922,
+      1.6380068063735962,
+      1.550830364227295,
+      1.5162874460220337,
+      1.5154277086257935,
+      1.5152742862701416,
+      1.510502576828003,
+      1.513155460357666,
+      1.513131856918335,
+      1.5095428228378296,
+      1.5045267343521118,
+      1.4966490268707275,
+      1.5024579763412476,
+      1.4976110458374023,
+      1.4977227449417114,
+      1.4980478286743164,
+      1.4940409660339355,
+      1.49163818359375,
+      1.4895493984222412,
+      1.4770550727844238,
+      1.4827024936676025,
+      1.4714136123657227,
+      1.4484158754348755,
+      1.4317924976348877,
+      1.3880096673965454,
+      1.354623794555664,
+      1.3511745929718018,
+      1.3489454984664917,
+      1.3342679738998413,
+      1.3305190801620483,
+      1.309164047241211,
+      1.2448537349700928,
+      1.2637850046157837,
+      1.2463005781173706,
+      1.2380549907684326,
+      1.2006921768188477,
+      1.1977745294570923,
+      1.2074775695800781,
+      1.141572117805481,
+      1.1027672290802002,
+      1.0749335289001465,
+      1.061755895614624,
+      1.0188356637954712,
+      1.0110366344451904,
+      1.0043277740478516,
+      1.0005559921264648,
+      0.9825484752655029,
+      0.9517877101898193,
+      0.9763262271881104,
+      0.9933456182479858,
+      0.959964394569397,
+      0.9830599427223206,
+      0.9790149927139282,
+      0.9470869898796082,
+      0.935240626335144,
+      0.9258978962898254,
+      0.9422906637191772,
+      0.9419942498207092,
+      0.9160555005073547
+    ],
+    "denoise_loss": [],
+    "ortho_loss": [
+      0.7092215418815613,
+      0.46743276715278625,
+      0.372333288192749,
+      0.3163504898548126,
+      0.29026153683662415,
+      0.2792986333370209,
+      0.27686527371406555,
+      0.262327641248703,
+      0.24570715427398682,
+      0.2475493848323822,
+      0.24840138852596283,
+      0.25525224208831787,
+      0.27400314807891846,
+      0.2834039628505707,
+      0.28120777010917664,
+      0.2767947316169739,
+      0.2838645875453949,
+      0.26357901096343994,
+      0.24860280752182007,
+      0.2444341778755188,
+      0.2471916675567627,
+      0.2405933439731598,
+      0.24468615651130676,
+      0.24482561647891998,
+      0.240675151348114,
+      0.24134141206741333,
+      0.24261823296546936,
+      0.24529020488262177,
+      0.24550136923789978,
+      0.24657639861106873,
+      0.24326443672180176,
+      0.24152472615242004,
+      0.24578434228897095,
+      0.24196375906467438,
+      0.24180641770362854,
+      0.24166396260261536,
+      0.23602785170078278,
+      0.2411782592535019,
+      0.23615865409374237,
+      0.23971572518348694,
+      0.23798082768917084,
+      0.2350902259349823,
+      0.23152871429920197,
+      0.23074424266815186,
+      0.22700950503349304,
+      0.2275175005197525,
+      0.23137938976287842,
+      0.22892071306705475,
+      0.22994598746299744,
+      0.2339673489332199,
+      0.23359638452529907,
+      0.23373164236545563,
+      0.23107723891735077,
+      0.2339753359556198,
+      0.23272015154361725,
+      0.2336493581533432,
+      0.233825221657753,
+      0.2339244931936264,
+      0.235467329621315,
+      0.2342061847448349,
+      0.23358942568302155,
+      0.23396462202072144,
+      0.23547381162643433,
+      0.23512768745422363,
+      0.23589524626731873,
+      0.2354237139225006,
+      0.2354665845632553,
+      0.23491325974464417,
+      0.23564963042736053,
+      0.2355274111032486
+    ],
+    "lr": [
+      1.6752136752136756e-05,
+      3.384615384615385e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      4e-05,
+      3.9947798576324814e-05,
+      3.8761402583706826e-05,
+      3.757500659108885e-05,
+      3.6388610598470864e-05,
+      3.5202214605852884e-05,
+      3.401581861323491e-05,
+      3.282942262061693e-05,
+      3.0670181914052204e-05,
+      2.948378592143422e-05,
+      2.8297389928816243e-05,
+      2.711099393619826e-05,
+      2.5924597943580284e-05,
+      2.4738201950962303e-05,
+      2.3551805958344316e-05,
+      2.1392565251779595e-05,
+      2.020616925916161e-05,
+      1.901977326654364e-05,
+      1.783337727392566e-05,
+      1.6646981281307675e-05,
+      1.546058528868969e-05,
+      1.427418929607171e-05,
+      1.2114948589506984e-05,
+      1.0928552596889013e-05,
+      9.742156604271029e-06,
+      8.555760611653046e-06,
+      7.369364619035064e-06,
+      6.182968626417082e-06,
+      4.996572633799099e-06
+    ],
+    "emb_lr": [],
+    "eval_step": [
+      350,
+      741,
+      1132,
+      1523,
+      1914,
+      2305,
+      2696,
+      3087,
+      3478,
+      3869
+    ],
+    "eval_accuracy": [
+      0.01,
+      0.0,
+      0.0,
+      0.0,
+      0.0,
+      0.0,
+      0.0,
+      0.0,
+      0.0,
+      0.0
+    ]
+  },
+  "final_accuracy": 0.28291666666666665,
+  "sft_eval": {
+    "config": {
+      "ops": "add_sub",
+      "K": null,
+      "mode": "sft",
+      "n_digits": 6,
+      "n_per_split": 100
+    },
+    "splits": {
+      "add_S0": {
+        "full_accuracy": 0.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 0.2066115702479339,
+            "count": 605
+          },
+          "SS": {
+            "accuracy": 1.0,
+            "count": 95
+          }
+        }
+      },
+      "add_S1": {
+        "full_accuracy": 0.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 0.2647058823529412,
+            "count": 204
+          },
+          "SC": {
+            "accuracy": 0.11242603550295859,
+            "count": 169
+          },
+          "SS": {
+            "accuracy": 0.6451612903225806,
+            "count": 31
+          },
+          "UC": {
+            "accuracy": 0.21621621621621623,
+            "count": 296
+          }
+        }
+      },
+      "add_S2": {
+        "full_accuracy": 0.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 0.3987730061349693,
+            "count": 163
+          },
+          "SC": {
+            "accuracy": 0.13846153846153847,
+            "count": 130
+          },
+          "SS": {
+            "accuracy": 0.3448275862068966,
+            "count": 87
+          },
+          "UC": {
+            "accuracy": 0.35960591133004927,
+            "count": 203
+          },
+          "US": {
+            "accuracy": 0.6324786324786325,
+            "count": 117
+          }
+        }
+      },
+      "add_S3": {
+        "full_accuracy": 0.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 0.48760330578512395,
+            "count": 121
+          },
+          "SC": {
+            "accuracy": 0.04132231404958678,
+            "count": 121
+          },
+          "SS": {
+            "accuracy": 0.6530612244897959,
+            "count": 49
+          },
+          "UC": {
+            "accuracy": 0.3548387096774194,
+            "count": 186
+          },
+          "US": {
+            "accuracy": 0.57847533632287,
+            "count": 223
+          }
+        }
+      },
+      "add_S4": {
+        "full_accuracy": 0.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 0.47115384615384615,
+            "count": 104
+          },
+          "SC": {
+            "accuracy": 0.10377358490566038,
+            "count": 106
+          },
+          "SS": {
+            "accuracy": 0.5652173913043478,
+            "count": 23
+          },
+          "UC": {
+            "accuracy": 0.3875,
+            "count": 160
+          },
+          "US": {
+            "accuracy": 0.5700325732899023,
+            "count": 307
+          }
+        }
+      },
+      "add_S5": {
+        "full_accuracy": 0.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 0.46,
+            "count": 100
+          },
+          "SC": {
+            "accuracy": 0.01,
+            "count": 100
+          },
+          "UC": {
+            "accuracy": 0.18,
+            "count": 100
+          },
+          "US": {
+            "accuracy": 0.09,
+            "count": 400
+          }
+        }
+      },
+      "add_S6": {
+        "full_accuracy": 0.02,
+        "n_examples": 100,
+        "per_subtask": {
+          "SC": {
+            "accuracy": 0.02,
+            "count": 100
+          },
+          "UC": {
+            "accuracy": 0.65,
+            "count": 100
+          },
+          "US": {
+            "accuracy": 0.648,
+            "count": 500
+          }
+        }
+      },
+      "add_random": {
+        "full_accuracy": 0.0,
+        "n_examples": 200,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 0.2639821029082774,
+            "count": 447
+          },
+          "SC": {
+            "accuracy": 0.1125,
+            "count": 320
+          },
+          "SS": {
+            "accuracy": 0.5178571428571429,
+            "count": 56
+          },
+          "UC": {
+            "accuracy": 0.23062381852551986,
+            "count": 529
+          },
+          "US": {
+            "accuracy": 0.5625,
+            "count": 48
+          }
+        }
+      },
+      "add_C3": {
+        "full_accuracy": 0.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 0.25,
+            "count": 300
+          },
+          "SC": {
+            "accuracy": 0.08,
+            "count": 100
+          },
+          "UC": {
+            "accuracy": 0.11398963730569948,
+            "count": 193
+          },
+          "US": {
+            "accuracy": 0.14953271028037382,
+            "count": 107
+          }
+        }
+      },
+      "add_C4": {
+        "full_accuracy": 0.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 0.355,
+            "count": 200
+          },
+          "SC": {
+            "accuracy": 0.03,
+            "count": 100
+          },
+          "UC": {
+            "accuracy": 0.10546875,
+            "count": 256
+          },
+          "US": {
+            "accuracy": 0.24305555555555555,
+            "count": 144
+          }
+        }
+      },
+      "add_C5": {
+        "full_accuracy": 0.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 0.5,
+            "count": 100
+          },
+          "SC": {
+            "accuracy": 0.09,
+            "count": 100
+          },
+          "UC": {
+            "accuracy": 0.13725490196078433,
+            "count": 306
+          },
+          "US": {
+            "accuracy": 0.3556701030927835,
+            "count": 194
+          }
+        }
+      },
+      "add_C6": {
+        "full_accuracy": 0.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "SC": {
+            "accuracy": 0.12,
+            "count": 100
+          },
+          "UC": {
+            "accuracy": 0.2540983606557377,
+            "count": 366
+          },
+          "US": {
+            "accuracy": 0.8632478632478633,
+            "count": 234
+          }
+        }
+      },
+      "sub_M0": {
+        "full_accuracy": 0.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 0.20465890183028287,
+            "count": 601
+          },
+          "ME": {
+            "accuracy": 1.0,
+            "count": 99
+          }
+        }
+      },
+      "sub_M1": {
+        "full_accuracy": 0.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 0.3835125448028674,
+            "count": 279
+          },
+          "MB": {
+            "accuracy": 0.0,
+            "count": 145
+          },
+          "ME": {
+            "accuracy": 1.0,
+            "count": 24
+          },
+          "UB": {
+            "accuracy": 0.09523809523809523,
+            "count": 252
+          }
+        }
+      },
+      "sub_M2": {
+        "full_accuracy": 0.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 0.6150234741784038,
+            "count": 213
+          },
+          "MB": {
+            "accuracy": 0.0,
+            "count": 113
+          },
+          "ME": {
+            "accuracy": 1.0,
+            "count": 85
+          },
+          "UB": {
+            "accuracy": 0.16574585635359115,
+            "count": 181
+          },
+          "UD": {
+            "accuracy": 0.0,
+            "count": 108
+          }
+        }
+      },
+      "sub_M3": {
+        "full_accuracy": 0.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 0.7597765363128491,
+            "count": 179
+          },
+          "MB": {
+            "accuracy": 0.0,
+            "count": 103
+          },
+          "ME": {
+            "accuracy": 1.0,
+            "count": 56
+          },
+          "UB": {
+            "accuracy": 0.12080536912751678,
+            "count": 149
+          },
+          "UD": {
+            "accuracy": 0.0,
+            "count": 213
+          }
+        }
+      },
+      "sub_M4": {
+        "full_accuracy": 0.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 0.5,
+            "count": 200
+          },
+          "MB": {
+            "accuracy": 0.0,
+            "count": 100
+          },
+          "UB": {
+            "accuracy": 0.3,
+            "count": 100
+          },
+          "UD": {
+            "accuracy": 0.0,
+            "count": 300
+          }
+        }
+      },
+      "sub_M5": {
+        "full_accuracy": 0.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "MB": {
+            "accuracy": 0.0,
+            "count": 100
+          },
+          "UB": {
+            "accuracy": 0.31,
+            "count": 100
+          },
+          "UD": {
+            "accuracy": 0.0,
+            "count": 400
+          }
+        }
+      },
+      "sub_random": {
+        "full_accuracy": 0.0,
+        "n_examples": 200,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 0.3616666666666667,
+            "count": 600
+          },
+          "MB": {
+            "accuracy": 0.0,
+            "count": 267
+          },
+          "ME": {
+            "accuracy": 1.0,
+            "count": 53
+          },
+          "UB": {
+            "accuracy": 0.12072892938496584,
+            "count": 439
+          },
+          "UD": {
+            "accuracy": 0.0,
+            "count": 41
+          }
+        }
+      },
+      "sub_B3": {
+        "full_accuracy": 0.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 0.3333333333333333,
+            "count": 300
+          },
+          "MB": {
+            "accuracy": 0.0,
+            "count": 100
+          },
+          "UB": {
+            "accuracy": 0.17766497461928935,
+            "count": 197
+          },
+          "UD": {
+            "accuracy": 0.0,
+            "count": 103
+          }
+        }
+      },
+      "sub_B4": {
+        "full_accuracy": 0.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 0.5,
+            "count": 200
+          },
+          "MB": {
+            "accuracy": 0.0,
+            "count": 100
+          },
+          "UB": {
+            "accuracy": 0.145748987854251,
+            "count": 247
+          },
+          "UD": {
+            "accuracy": 0.0,
+            "count": 153
+          }
+        }
+      },
+      "sub_B5": {
+        "full_accuracy": 0.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "MB": {
+            "accuracy": 0.0,
+            "count": 100
+          },
+          "UB": {
+            "accuracy": 0.11073825503355705,
+            "count": 298
+          },
+          "UD": {
+            "accuracy": 0.0,
+            "count": 202
+          }
+        }
+      }
+    },
+    "summary": {
+      "overall_accuracy": 0.0008333333333333334,
+      "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": 0.74,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 0.9603305785123967,
+            "count": 605
+          },
+          "SS": {
+            "accuracy": 0.9578947368421052,
+            "count": 95
+          }
+        }
+      },
+      "add_S1": {
+        "full_accuracy": 0.58,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 0.9362745098039216,
+            "count": 204
+          },
+          "SC": {
+            "accuracy": 0.9585798816568047,
+            "count": 169
+          },
+          "SS": {
+            "accuracy": 1.0,
+            "count": 31
+          },
+          "UC": {
+            "accuracy": 0.875,
+            "count": 296
+          }
+        }
+      },
+      "add_S2": {
+        "full_accuracy": 0.35,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 0.9693251533742331,
+            "count": 163
+          },
+          "SC": {
+            "accuracy": 0.9076923076923077,
+            "count": 130
+          },
+          "SS": {
+            "accuracy": 0.9080459770114943,
+            "count": 87
+          },
+          "UC": {
+            "accuracy": 0.7192118226600985,
+            "count": 203
+          },
+          "US": {
+            "accuracy": 0.9316239316239316,
+            "count": 117
+          }
+        }
+      },
+      "add_S3": {
+        "full_accuracy": 0.19,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 0.9834710743801653,
+            "count": 121
+          },
+          "SC": {
+            "accuracy": 0.9504132231404959,
+            "count": 121
+          },
+          "SS": {
+            "accuracy": 0.9183673469387755,
+            "count": 49
+          },
+          "UC": {
+            "accuracy": 0.6612903225806451,
+            "count": 186
+          },
+          "US": {
+            "accuracy": 0.6860986547085202,
+            "count": 223
+          }
+        }
+      },
+      "add_S4": {
+        "full_accuracy": 0.2,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 104
+          },
+          "SC": {
+            "accuracy": 0.9433962264150944,
+            "count": 106
+          },
+          "SS": {
+            "accuracy": 1.0,
+            "count": 23
+          },
+          "UC": {
+            "accuracy": 0.64375,
+            "count": 160
+          },
+          "US": {
+            "accuracy": 0.5309446254071661,
+            "count": 307
+          }
+        }
+      },
+      "add_S5": {
+        "full_accuracy": 0.05,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "SC": {
+            "accuracy": 0.98,
+            "count": 100
+          },
+          "UC": {
+            "accuracy": 0.36,
+            "count": 100
+          },
+          "US": {
+            "accuracy": 0.3325,
+            "count": 400
+          }
+        }
+      },
+      "add_S6": {
+        "full_accuracy": 0.03,
+        "n_examples": 100,
+        "per_subtask": {
+          "SC": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "UC": {
+            "accuracy": 0.28,
+            "count": 100
+          },
+          "US": {
+            "accuracy": 0.298,
+            "count": 500
+          }
+        }
+      },
+      "add_random": {
+        "full_accuracy": 0.575,
+        "n_examples": 200,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 0.9552572706935123,
+            "count": 447
+          },
+          "SC": {
+            "accuracy": 0.9625,
+            "count": 320
+          },
+          "SS": {
+            "accuracy": 0.9821428571428571,
+            "count": 56
+          },
+          "UC": {
+            "accuracy": 0.8657844990548205,
+            "count": 529
+          },
+          "US": {
+            "accuracy": 0.7916666666666666,
+            "count": 48
+          }
+        }
+      },
+      "add_C3": {
+        "full_accuracy": 0.41,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 0.9833333333333333,
+            "count": 300
+          },
+          "SC": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "UC": {
+            "accuracy": 0.7150259067357513,
+            "count": 193
+          },
+          "US": {
+            "accuracy": 0.7102803738317757,
+            "count": 107
+          }
+        }
+      },
+      "add_C4": {
+        "full_accuracy": 0.32,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 200
+          },
+          "SC": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "UC": {
+            "accuracy": 0.71875,
+            "count": 256
+          },
+          "US": {
+            "accuracy": 0.7291666666666666,
+            "count": 144
+          }
+        }
+      },
+      "add_C5": {
+        "full_accuracy": 0.23,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "SC": {
+            "accuracy": 0.97,
+            "count": 100
+          },
+          "UC": {
+            "accuracy": 0.696078431372549,
+            "count": 306
+          },
+          "US": {
+            "accuracy": 0.6752577319587629,
+            "count": 194
+          }
+        }
+      },
+      "add_C6": {
+        "full_accuracy": 0.23,
+        "n_examples": 100,
+        "per_subtask": {
+          "SC": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "UC": {
+            "accuracy": 0.7486338797814208,
+            "count": 366
+          },
+          "US": {
+            "accuracy": 0.7136752136752137,
+            "count": 234
+          }
+        }
+      },
+      "sub_M0": {
+        "full_accuracy": 0.61,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 0.9234608985024958,
+            "count": 601
+          },
+          "ME": {
+            "accuracy": 0.9797979797979798,
+            "count": 99
+          }
+        }
+      },
+      "sub_M1": {
+        "full_accuracy": 0.45,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 0.9605734767025089,
+            "count": 279
+          },
+          "MB": {
+            "accuracy": 0.9241379310344827,
+            "count": 145
+          },
+          "ME": {
+            "accuracy": 0.875,
+            "count": 24
+          },
+          "UB": {
+            "accuracy": 0.8015873015873016,
+            "count": 252
+          }
+        }
+      },
+      "sub_M2": {
+        "full_accuracy": 0.21,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 0.9530516431924883,
+            "count": 213
+          },
+          "MB": {
+            "accuracy": 0.9380530973451328,
+            "count": 113
+          },
+          "ME": {
+            "accuracy": 0.9764705882352941,
+            "count": 85
+          },
+          "UB": {
+            "accuracy": 0.5524861878453039,
+            "count": 181
+          },
+          "UD": {
+            "accuracy": 0.7685185185185185,
+            "count": 108
+          }
+        }
+      },
+      "sub_M3": {
+        "full_accuracy": 0.07,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 0.9888268156424581,
+            "count": 179
+          },
+          "MB": {
+            "accuracy": 0.941747572815534,
+            "count": 103
+          },
+          "ME": {
+            "accuracy": 1.0,
+            "count": 56
+          },
+          "UB": {
+            "accuracy": 0.4563758389261745,
+            "count": 149
+          },
+          "UD": {
+            "accuracy": 0.4788732394366197,
+            "count": 213
+          }
+        }
+      },
+      "sub_M4": {
+        "full_accuracy": 0.03,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 0.935,
+            "count": 200
+          },
+          "MB": {
+            "accuracy": 0.96,
+            "count": 100
+          },
+          "UB": {
+            "accuracy": 0.34,
+            "count": 100
+          },
+          "UD": {
+            "accuracy": 0.23333333333333334,
+            "count": 300
+          }
+        }
+      },
+      "sub_M5": {
+        "full_accuracy": 0.06,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "MB": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "UB": {
+            "accuracy": 0.38,
+            "count": 100
+          },
+          "UD": {
+            "accuracy": 0.235,
+            "count": 400
+          }
+        }
+      },
+      "sub_random": {
+        "full_accuracy": 0.32,
+        "n_examples": 200,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 0.935,
+            "count": 600
+          },
+          "MB": {
+            "accuracy": 0.898876404494382,
+            "count": 267
+          },
+          "ME": {
+            "accuracy": 0.9433962264150944,
+            "count": 53
+          },
+          "UB": {
+            "accuracy": 0.7562642369020501,
+            "count": 439
+          },
+          "UD": {
+            "accuracy": 0.7804878048780488,
+            "count": 41
+          }
+        }
+      },
+      "sub_B3": {
+        "full_accuracy": 0.13,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 0.8833333333333333,
+            "count": 300
+          },
+          "MB": {
+            "accuracy": 0.95,
+            "count": 100
+          },
+          "UB": {
+            "accuracy": 0.5482233502538071,
+            "count": 197
+          },
+          "UD": {
+            "accuracy": 0.5533980582524272,
+            "count": 103
+          }
+        }
+      },
+      "sub_B4": {
+        "full_accuracy": 0.09,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 0.95,
+            "count": 200
+          },
+          "MB": {
+            "accuracy": 0.97,
+            "count": 100
+          },
+          "UB": {
+            "accuracy": 0.5546558704453441,
+            "count": 247
+          },
+          "UD": {
+            "accuracy": 0.47058823529411764,
+            "count": 153
+          }
+        }
+      },
+      "sub_B5": {
+        "full_accuracy": 0.04,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "MB": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "UB": {
+            "accuracy": 0.540268456375839,
+            "count": 298
+          },
+          "UD": {
+            "accuracy": 0.4603960396039604,
+            "count": 202
+          }
+        }
+      }
+    },
+    "summary": {
+      "overall_accuracy": 0.28291666666666665,
+      "total_examples": 2400,
+      "n_splits": 22
+    }
+  },
+  "sorl_overall_accuracy": 0.28291666666666665,
+  "sft_overall_accuracy": 0.0008333333333333334
+}

add_sub_sorl_v1_abs10_K1_25K_1L3H510d/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:83dc63f41a380d62231535ea7934fea7ce172db8bc245f6beae0be7d7b4c8b71
+size 634679036

add_sub_sorl_v1_abs10_K1_25K_1L3H510d/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": 4e-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_1L3H510d",
+  "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": 3,
+  "n_embd": 510,
+  "ops": "add_sub",
+  "abs_vocab": 10,
+  "dataset_size": 25000,
+  "mode": "sorl",
+  "device": "cuda",
+  "push_to_hub": true,
+  "no_wandb": false,
+  "n_params": 158593426,
+  "run_name": "add_sub_sorl_v1_abs10_K1_25K_1L3H510d",
+  "git_commit": "f447da529caceac8c7d256cbb2cd185cbc50feac",
+  "timestamp": "2026-04-12T16:20:24.628526+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": "ybv7051n",
+  "wandb_url": "https://wandb.ai/nlp_and_interpretability/sorl-arithmetic/runs/ybv7051n",
+  "final_accuracy": 0.28291666666666665,
+  "sft_accuracy": 0.0008333333333333334,
+  "eval_method": "ArithmeticEvaluator"
+}