Upload add_sub_sorl_v1_abs30_K1_10K

add_sub_sorl_v1_abs30_K1_10K/config.json

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

add_sub_sorl_v1_abs30_K1_10K/generation_config.json

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

add_sub_sorl_v1_abs30_K1_10K/metrics.json

@@ -0,0 +1,1557 @@
+{
+  "history": {
+    "step": [
+      50,
+      100,
+      150,
+      207,
+      257,
+      307,
+      364,
+      414,
+      464,
+      521,
+      571,
+      621,
+      678,
+      728,
+      778,
+      835,
+      885,
+      935,
+      992,
+      1042,
+      1092,
+      1149,
+      1199,
+      1249,
+      1306,
+      1356,
+      1406,
+      1463,
+      1513,
+      1563,
+      1620,
+      1670,
+      1720,
+      1777,
+      1827,
+      1877,
+      1934,
+      1984,
+      2034,
+      2091,
+      2141,
+      2191,
+      2248,
+      2298,
+      2348,
+      2405,
+      2455,
+      2505,
+      2562,
+      2612,
+      2662,
+      2719,
+      2769,
+      2819,
+      2876,
+      2926,
+      2976,
+      3033,
+      3083,
+      3133
+    ],
+    "loss": [
+      7.419157981872559,
+      4.167179584503174,
+      2.757251739501953,
+      2.7242889404296875,
+      2.4392311573028564,
+      2.12624454498291,
+      1.7021116018295288,
+      0.1742163896560669,
+      -4.171255588531494,
+      -7.291606426239014,
+      -7.83780574798584,
+      -4.589130401611328,
+      -2.949950695037842,
+      -1.528172254562378,
+      -0.9651578664779663,
+      -0.8742300271987915,
+      -0.550557017326355,
+      -0.38707196712493896,
+      -0.5141674876213074,
+      -0.6698199510574341,
+      -0.7084141373634338,
+      -0.47245967388153076,
+      -0.31419694423675537,
+      -0.31762227416038513,
+      -0.24917703866958618,
+      -0.5615812540054321,
+      -0.11299914121627808,
+      -0.3741247057914734,
+      -0.1392630785703659,
+      -0.4355674088001251,
+      -0.7360162138938904,
+      -0.5835363268852234,
+      -0.5749945640563965,
+      -0.4588763415813446,
+      -0.5538105964660645,
+      -0.009306885302066803,
+      -0.4344196915626526,
+      -0.23470139503479004,
+      -0.4863656759262085,
+      -0.6077032685279846,
+      -0.41746845841407776,
+      -0.4240177273750305,
+      -0.17653952538967133,
+      -0.39587831497192383,
+      -0.1691063493490219,
+      -0.12267603725194931,
+      -0.3334910571575165,
+      -0.11227956414222717,
+      -0.2049349546432495,
+      -0.3078603148460388,
+      -0.17479164898395538,
+      -0.25693076848983765,
+      -0.22230546176433563,
+      -0.24037127196788788,
+      -0.07592999190092087,
+      -0.35476192831993103,
+      -0.3569074273109436,
+      -0.22312912344932556,
+      -0.22871972620487213,
+      -0.16561374068260193
+    ],
+    "base_loss": [
+      7.4173455238342285,
+      3.6480441093444824,
+      1.9713058471679688,
+      1.8219451904296875,
+      1.865543007850647,
+      1.8542261123657227,
+      1.805886149406433,
+      1.7817356586456299,
+      1.8499161005020142,
+      1.6429988145828247,
+      1.4622584581375122,
+      0.8687496185302734,
+      0.5731961131095886,
+      0.35015660524368286,
+      0.2690192759037018,
+      0.22296881675720215,
+      0.1571018397808075,
+      0.16091905534267426,
+      0.1413668543100357,
+      0.14242176711559296,
+      0.15262103080749512,
+      0.11749549955129623,
+      0.1065789982676506,
+      0.08413096517324448,
+      0.07932066917419434,
+      0.11860263347625732,
+      0.07409992069005966,
+      0.07803652435541153,
+      0.047815799713134766,
+      0.10229728370904922,
+      0.11566775292158127,
+      0.10984303057193756,
+      0.09876441955566406,
+      0.0688941702246666,
+      0.08261072635650635,
+      0.053319063037633896,
+      0.06365398317575455,
+      0.0505535863339901,
+      0.06858514994382858,
+      0.07909546047449112,
+      0.060411665588617325,
+      0.0625351145863533,
+      0.032719288021326065,
+      0.05432973429560661,
+      0.045225996524095535,
+      0.021227678284049034,
+      0.050432708114385605,
+      0.027083681896328926,
+      0.031196508556604385,
+      0.04615318775177002,
+      0.027639741078019142,
+      0.05231175199151039,
+      0.03158637136220932,
+      0.032559193670749664,
+      0.01280138734728098,
+      0.04558839276432991,
+      0.04613020643591881,
+      0.03275144845247269,
+      0.03132258728146553,
+      0.024606991559267044
+    ],
+    "info_loss": [
+      -0.5185031890869141,
+      -0.09349203109741211,
+      -0.03490471839904785,
+      -0.02159249782562256,
+      -0.054520368576049805,
+      -0.08344614505767822,
+      -0.1214907169342041,
+      -0.27150094509124756,
+      -0.7121474742889404,
+      -1.0041959285736084,
+      -1.0389212369918823,
+      -0.6523160934448242,
+      -0.45543068647384644,
+      -0.28470802307128906,
+      -0.21153677999973297,
+      -0.1875189244747162,
+      -0.14062973856925964,
+      -0.12236915528774261,
+      -0.12718062102794647,
+      -0.1365828663110733,
+      -0.13193942606449127,
+      -0.10466521978378296,
+      -0.08482377231121063,
+      -0.0746171772480011,
+      -0.07218538224697113,
+      -0.10219592601060867,
+      -0.04982271045446396,
+      -0.07324571162462234,
+      -0.038586750626564026,
+      -0.07255075871944427,
+      -0.10315456241369247,
+      -0.08851015567779541,
+      -0.08371994644403458,
+      -0.06724002212285995,
+      -0.0802496150135994,
+      -0.021388381719589233,
+      -0.06261564046144485,
+      -0.0424460805952549,
+      -0.06758096814155579,
+      -0.07665054500102997,
+      -0.058085739612579346,
+      -0.05829145759344101,
+      -0.030241522938013077,
+      -0.053621694445610046,
+      -0.03214117884635925,
+      -0.020935092121362686,
+      -0.04635276272892952,
+      -0.02176225744187832,
+      -0.03096679225564003,
+      -0.04063677042722702,
+      -0.02714877389371395,
+      -0.036504216492176056,
+      -0.031483057886362076,
+      -0.032049935311079025,
+      -0.012561824172735214,
+      -0.045204248279333115,
+      -0.04523665830492973,
+      -0.0305631123483181,
+      -0.0310970451682806,
+      -0.023466313257813454
+    ],
+    "abs_loss": [
+      3.381772994995117,
+      2.774198293685913,
+      2.777165412902832,
+      2.7333767414093018,
+      2.7525675296783447,
+      2.701953649520874,
+      2.742412567138672,
+      2.6781795024871826,
+      2.65073823928833,
+      2.6559343338012695,
+      2.530367612838745,
+      2.4332215785980225,
+      2.2735555171966553,
+      1.9705966711044312,
+      1.622453212738037,
+      1.4155868291854858,
+      1.1430540084838867,
+      0.9845269322395325,
+      0.9937983155250549,
+      0.802861750125885,
+      0.7566664814949036,
+      0.6577289700508118,
+      0.7418577075004578,
+      0.5929294228553772,
+      0.5775941014289856,
+      0.5590793490409851,
+      0.6108124256134033,
+      0.4465348720550537,
+      0.393932968378067,
+      0.3919040262699127,
+      0.41778573393821716,
+      0.4128030240535736,
+      0.36697685718536377,
+      0.3254868686199188,
+      0.3693109452724457,
+      0.3870414197444916,
+      0.2671542167663574,
+      0.23766420781612396,
+      0.27502405643463135,
+      0.19808150827884674,
+      0.2262984663248062,
+      0.24667154252529144,
+      0.24236536026000977,
+      0.20162977278232574,
+      0.2073269635438919,
+      0.18350571393966675,
+      0.1846071034669876,
+      0.15442465245723724,
+      0.1817246526479721,
+      0.14412124454975128,
+      0.14306403696537018,
+      0.12268459051847458,
+      0.13687384128570557,
+      0.11778446286916733,
+      0.11796697229146957,
+      0.10396292060613632,
+      0.11473483592271805,
+      0.10343429446220398,
+      0.07652456313371658,
+      0.10590269416570663
+    ],
+    "zipf_loss": [
+      4.848667144775391,
+      1.176635980606079,
+      0.8572765588760376,
+      0.8449309468269348,
+      0.8436350226402283,
+      0.8362844586372375,
+      0.8368913531303406,
+      0.8396722078323364,
+      0.8352292776107788,
+      0.8417611122131348,
+      0.836111843585968,
+      0.821958601474762,
+      0.8038046956062317,
+      0.7716918587684631,
+      0.7189455032348633,
+      0.6364318132400513,
+      0.5843331813812256,
+      0.5772478580474854,
+      0.51689213514328,
+      0.4733008146286011,
+      0.3826923966407776,
+      0.3909241557121277,
+      0.35327601432800293,
+      0.28512564301490784,
+      0.3355967402458191,
+      0.28586745262145996,
+      0.25004681944847107,
+      0.23564238846302032,
+      0.15939532220363617,
+      0.148452490568161,
+      0.13808311522006989,
+      0.15044189989566803,
+      0.12674278020858765,
+      0.11208108067512512,
+      0.12914372980594635,
+      0.11255372315645218,
+      0.10136730968952179,
+      0.11543937772512436,
+      0.09335644543170929,
+      0.05989854037761688,
+      0.08034739643335342,
+      0.07169462740421295,
+      0.06891988217830658,
+      0.06584595888853073,
+      0.08634673058986664,
+      0.04709663242101669,
+      0.061143141239881516,
+      0.06281687319278717,
+      0.055363979190588,
+      0.037942074239254,
+      0.054749950766563416,
+      0.04353116825222969,
+      0.04725135862827301,
+      0.03579045459628105,
+      0.02509017288684845,
+      0.041295863687992096,
+      0.03785546123981476,
+      0.03940713033080101,
+      0.04327566549181938,
+      0.03385213017463684
+    ],
+    "denoise_loss": [],
+    "ortho_loss": [
+      0.3230772912502289,
+      0.1277298629283905,
+      0.10143587738275528,
+      0.08668191730976105,
+      0.09006865322589874,
+      0.07957521826028824,
+      0.09151509404182434,
+      0.09135851263999939,
+      0.1040191501379013,
+      0.118146151304245,
+      0.12644976377487183,
+      0.13958248496055603,
+      0.15548494458198547,
+      0.16445577144622803,
+      0.17663978040218353,
+      0.17986539006233215,
+      0.1787496656179428,
+      0.18172365427017212,
+      0.184878870844841,
+      0.18665480613708496,
+      0.18443602323532104,
+      0.17972944676876068,
+      0.1847168207168579,
+      0.18398721516132355,
+      0.18395553529262543,
+      0.1859062761068344,
+      0.1844593584537506,
+      0.19130879640579224,
+      0.18957658112049103,
+      0.18175643682479858,
+      0.18003104627132416,
+      0.18403305113315582,
+      0.1801728755235672,
+      0.17507442831993103,
+      0.18054728209972382,
+      0.17856059968471527,
+      0.1795228123664856,
+      0.17617978155612946,
+      0.17519906163215637,
+      0.17287763953208923,
+      0.1746804416179657,
+      0.1751648634672165,
+      0.17409953474998474,
+      0.17315584421157837,
+      0.17171329259872437,
+      0.17051003873348236,
+      0.17108182609081268,
+      0.1729588657617569,
+      0.16980983316898346,
+      0.1726285070180893,
+      0.1682920753955841,
+      0.1723119467496872,
+      0.1724519282579422,
+      0.17354536056518555,
+      0.17300596833229065,
+      0.17255157232284546,
+      0.17369846999645233,
+      0.17459796369075775,
+      0.17422601580619812,
+      0.17440278828144073
+    ],
+    "lr": [
+      3.9200000000000004e-05,
+      7.92e-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,
+      7.946710526315791e-05,
+      7.650657894736843e-05,
+      7.354605263157895e-05,
+      7.017105263157896e-05,
+      6.721052631578948e-05,
+      6.425e-05,
+      6.0875e-05,
+      5.791447368421054e-05,
+      5.495394736842105e-05,
+      5.157894736842105e-05,
+      4.861842105263157e-05,
+      4.565789473684212e-05,
+      4.2282894736842104e-05,
+      3.9322368421052625e-05,
+      3.636184210526315e-05,
+      3.2986842105263165e-05,
+      3.0026315789473686e-05,
+      2.7065789473684206e-05,
+      2.3690789473684223e-05,
+      2.0730263157894743e-05,
+      1.7769736842105264e-05,
+      1.4394736842105275e-05,
+      1.1434210526315796e-05,
+      8.473684210526318e-06
+    ],
+    "emb_lr": [],
+    "eval_step": [
+      150,
+      307,
+      464,
+      621,
+      778,
+      935,
+      1092,
+      1199,
+      1356,
+      1513,
+      1670,
+      1827,
+      1984,
+      2141,
+      2298,
+      2455,
+      2612,
+      2769,
+      2926,
+      3083
+    ],
+    "eval_accuracy": [
+      0.01,
+      0.0,
+      0.0,
+      0.0,
+      0.0,
+      0.0,
+      0.0,
+      0.0,
+      0.0,
+      0.0,
+      0.0,
+      0.0,
+      0.0,
+      0.0,
+      0.0,
+      0.0,
+      0.0,
+      0.0,
+      0.0,
+      0.0
+    ]
+  },
+  "final_accuracy": 0.96125,
+  "sft_eval": {
+    "config": {
+      "ops": "add_sub",
+      "K": null,
+      "mode": "sft",
+      "n_digits": 6,
+      "n_per_split": 100
+    },
+    "splits": {
+      "add_S0": {
+        "full_accuracy": 0.97,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 0.9950413223140496,
+            "count": 605
+          },
+          "SS": {
+            "accuracy": 1.0,
+            "count": 95
+          }
+        }
+      },
+      "add_S1": {
+        "full_accuracy": 1.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 204
+          },
+          "SC": {
+            "accuracy": 1.0,
+            "count": 169
+          },
+          "SS": {
+            "accuracy": 1.0,
+            "count": 31
+          },
+          "UC": {
+            "accuracy": 1.0,
+            "count": 296
+          }
+        }
+      },
+      "add_S2": {
+        "full_accuracy": 0.96,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 163
+          },
+          "SC": {
+            "accuracy": 0.9769230769230769,
+            "count": 130
+          },
+          "SS": {
+            "accuracy": 0.9770114942528736,
+            "count": 87
+          },
+          "UC": {
+            "accuracy": 1.0,
+            "count": 203
+          },
+          "US": {
+            "accuracy": 1.0,
+            "count": 117
+          }
+        }
+      },
+      "add_S3": {
+        "full_accuracy": 0.64,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 121
+          },
+          "SC": {
+            "accuracy": 0.9834710743801653,
+            "count": 121
+          },
+          "SS": {
+            "accuracy": 1.0,
+            "count": 49
+          },
+          "UC": {
+            "accuracy": 0.8172043010752689,
+            "count": 186
+          },
+          "US": {
+            "accuracy": 1.0,
+            "count": 223
+          }
+        }
+      },
+      "add_S4": {
+        "full_accuracy": 0.59,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 104
+          },
+          "SC": {
+            "accuracy": 1.0,
+            "count": 106
+          },
+          "SS": {
+            "accuracy": 1.0,
+            "count": 23
+          },
+          "UC": {
+            "accuracy": 0.8125,
+            "count": 160
+          },
+          "US": {
+            "accuracy": 0.9153094462540716,
+            "count": 307
+          }
+        }
+      },
+      "add_S5": {
+        "full_accuracy": 0.62,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "SC": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "UC": {
+            "accuracy": 0.71,
+            "count": 100
+          },
+          "US": {
+            "accuracy": 0.86,
+            "count": 400
+          }
+        }
+      },
+      "add_S6": {
+        "full_accuracy": 1.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "SC": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "UC": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "US": {
+            "accuracy": 1.0,
+            "count": 500
+          }
+        }
+      },
+      "add_random": {
+        "full_accuracy": 0.99,
+        "n_examples": 200,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 447
+          },
+          "SC": {
+            "accuracy": 0.996875,
+            "count": 320
+          },
+          "SS": {
+            "accuracy": 1.0,
+            "count": 56
+          },
+          "UC": {
+            "accuracy": 0.998109640831758,
+            "count": 529
+          },
+          "US": {
+            "accuracy": 1.0,
+            "count": 48
+          }
+        }
+      },
+      "add_C3": {
+        "full_accuracy": 0.84,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 300
+          },
+          "SC": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "UC": {
+            "accuracy": 0.917098445595855,
+            "count": 193
+          },
+          "US": {
+            "accuracy": 1.0,
+            "count": 107
+          }
+        }
+      },
+      "add_C4": {
+        "full_accuracy": 0.78,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 200
+          },
+          "SC": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "UC": {
+            "accuracy": 0.92578125,
+            "count": 256
+          },
+          "US": {
+            "accuracy": 0.9583333333333334,
+            "count": 144
+          }
+        }
+      },
+      "add_C5": {
+        "full_accuracy": 0.72,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "SC": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "UC": {
+            "accuracy": 0.9313725490196079,
+            "count": 306
+          },
+          "US": {
+            "accuracy": 0.9536082474226805,
+            "count": 194
+          }
+        }
+      },
+      "add_C6": {
+        "full_accuracy": 0.82,
+        "n_examples": 100,
+        "per_subtask": {
+          "SC": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "UC": {
+            "accuracy": 0.953551912568306,
+            "count": 366
+          },
+          "US": {
+            "accuracy": 0.9957264957264957,
+            "count": 234
+          }
+        }
+      },
+      "sub_M0": {
+        "full_accuracy": 0.96,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 0.9933444259567388,
+            "count": 601
+          },
+          "ME": {
+            "accuracy": 1.0,
+            "count": 99
+          }
+        }
+      },
+      "sub_M1": {
+        "full_accuracy": 1.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 279
+          },
+          "MB": {
+            "accuracy": 1.0,
+            "count": 145
+          },
+          "ME": {
+            "accuracy": 1.0,
+            "count": 24
+          },
+          "UB": {
+            "accuracy": 1.0,
+            "count": 252
+          }
+        }
+      },
+      "sub_M2": {
+        "full_accuracy": 1.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 213
+          },
+          "MB": {
+            "accuracy": 1.0,
+            "count": 113
+          },
+          "ME": {
+            "accuracy": 1.0,
+            "count": 85
+          },
+          "UB": {
+            "accuracy": 1.0,
+            "count": 181
+          },
+          "UD": {
+            "accuracy": 1.0,
+            "count": 108
+          }
+        }
+      },
+      "sub_M3": {
+        "full_accuracy": 0.35,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 179
+          },
+          "MB": {
+            "accuracy": 1.0,
+            "count": 103
+          },
+          "ME": {
+            "accuracy": 1.0,
+            "count": 56
+          },
+          "UB": {
+            "accuracy": 0.5637583892617449,
+            "count": 149
+          },
+          "UD": {
+            "accuracy": 1.0,
+            "count": 213
+          }
+        }
+      },
+      "sub_M4": {
+        "full_accuracy": 0.03,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 200
+          },
+          "MB": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "UB": {
+            "accuracy": 0.49,
+            "count": 100
+          },
+          "UD": {
+            "accuracy": 0.6766666666666666,
+            "count": 300
+          }
+        }
+      },
+      "sub_M5": {
+        "full_accuracy": 0.0,
+        "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": 0.4975,
+            "count": 400
+          }
+        }
+      },
+      "sub_random": {
+        "full_accuracy": 0.995,
+        "n_examples": 200,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 600
+          },
+          "MB": {
+            "accuracy": 1.0,
+            "count": 267
+          },
+          "ME": {
+            "accuracy": 1.0,
+            "count": 53
+          },
+          "UB": {
+            "accuracy": 0.9977220956719818,
+            "count": 439
+          },
+          "UD": {
+            "accuracy": 1.0,
+            "count": 41
+          }
+        }
+      },
+      "sub_B3": {
+        "full_accuracy": 0.84,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 300
+          },
+          "MB": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "UB": {
+            "accuracy": 0.9187817258883249,
+            "count": 197
+          },
+          "UD": {
+            "accuracy": 1.0,
+            "count": 103
+          }
+        }
+      },
+      "sub_B4": {
+        "full_accuracy": 0.76,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 200
+          },
+          "MB": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "UB": {
+            "accuracy": 0.9311740890688259,
+            "count": 247
+          },
+          "UD": {
+            "accuracy": 0.9084967320261438,
+            "count": 153
+          }
+        }
+      },
+      "sub_B5": {
+        "full_accuracy": 0.59,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "MB": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "UB": {
+            "accuracy": 0.8859060402684564,
+            "count": 298
+          },
+          "UD": {
+            "accuracy": 0.8861386138613861,
+            "count": 202
+          }
+        }
+      }
+    },
+    "summary": {
+      "overall_accuracy": 0.7683333333333333,
+      "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": 1.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 605
+          },
+          "SS": {
+            "accuracy": 1.0,
+            "count": 95
+          }
+        }
+      },
+      "add_S1": {
+        "full_accuracy": 1.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 204
+          },
+          "SC": {
+            "accuracy": 1.0,
+            "count": 169
+          },
+          "SS": {
+            "accuracy": 1.0,
+            "count": 31
+          },
+          "UC": {
+            "accuracy": 1.0,
+            "count": 296
+          }
+        }
+      },
+      "add_S2": {
+        "full_accuracy": 1.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 163
+          },
+          "SC": {
+            "accuracy": 1.0,
+            "count": 130
+          },
+          "SS": {
+            "accuracy": 1.0,
+            "count": 87
+          },
+          "UC": {
+            "accuracy": 1.0,
+            "count": 203
+          },
+          "US": {
+            "accuracy": 1.0,
+            "count": 117
+          }
+        }
+      },
+      "add_S3": {
+        "full_accuracy": 1.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 121
+          },
+          "SC": {
+            "accuracy": 1.0,
+            "count": 121
+          },
+          "SS": {
+            "accuracy": 1.0,
+            "count": 49
+          },
+          "UC": {
+            "accuracy": 1.0,
+            "count": 186
+          },
+          "US": {
+            "accuracy": 1.0,
+            "count": 223
+          }
+        }
+      },
+      "add_S4": {
+        "full_accuracy": 1.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 104
+          },
+          "SC": {
+            "accuracy": 1.0,
+            "count": 106
+          },
+          "SS": {
+            "accuracy": 1.0,
+            "count": 23
+          },
+          "UC": {
+            "accuracy": 1.0,
+            "count": 160
+          },
+          "US": {
+            "accuracy": 1.0,
+            "count": 307
+          }
+        }
+      },
+      "add_S5": {
+        "full_accuracy": 0.99,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "SC": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "UC": {
+            "accuracy": 0.99,
+            "count": 100
+          },
+          "US": {
+            "accuracy": 1.0,
+            "count": 400
+          }
+        }
+      },
+      "add_S6": {
+        "full_accuracy": 1.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "SC": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "UC": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "US": {
+            "accuracy": 1.0,
+            "count": 500
+          }
+        }
+      },
+      "add_random": {
+        "full_accuracy": 1.0,
+        "n_examples": 200,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 447
+          },
+          "SC": {
+            "accuracy": 1.0,
+            "count": 320
+          },
+          "SS": {
+            "accuracy": 1.0,
+            "count": 56
+          },
+          "UC": {
+            "accuracy": 1.0,
+            "count": 529
+          },
+          "US": {
+            "accuracy": 1.0,
+            "count": 48
+          }
+        }
+      },
+      "add_C3": {
+        "full_accuracy": 1.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 300
+          },
+          "SC": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "UC": {
+            "accuracy": 1.0,
+            "count": 193
+          },
+          "US": {
+            "accuracy": 1.0,
+            "count": 107
+          }
+        }
+      },
+      "add_C4": {
+        "full_accuracy": 1.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 200
+          },
+          "SC": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "UC": {
+            "accuracy": 1.0,
+            "count": 256
+          },
+          "US": {
+            "accuracy": 1.0,
+            "count": 144
+          }
+        }
+      },
+      "add_C5": {
+        "full_accuracy": 1.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "SC": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "UC": {
+            "accuracy": 1.0,
+            "count": 306
+          },
+          "US": {
+            "accuracy": 1.0,
+            "count": 194
+          }
+        }
+      },
+      "add_C6": {
+        "full_accuracy": 1.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "SC": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "UC": {
+            "accuracy": 1.0,
+            "count": 366
+          },
+          "US": {
+            "accuracy": 1.0,
+            "count": 234
+          }
+        }
+      },
+      "sub_M0": {
+        "full_accuracy": 1.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 601
+          },
+          "ME": {
+            "accuracy": 1.0,
+            "count": 99
+          }
+        }
+      },
+      "sub_M1": {
+        "full_accuracy": 1.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 279
+          },
+          "MB": {
+            "accuracy": 1.0,
+            "count": 145
+          },
+          "ME": {
+            "accuracy": 1.0,
+            "count": 24
+          },
+          "UB": {
+            "accuracy": 1.0,
+            "count": 252
+          }
+        }
+      },
+      "sub_M2": {
+        "full_accuracy": 1.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 213
+          },
+          "MB": {
+            "accuracy": 1.0,
+            "count": 113
+          },
+          "ME": {
+            "accuracy": 1.0,
+            "count": 85
+          },
+          "UB": {
+            "accuracy": 1.0,
+            "count": 181
+          },
+          "UD": {
+            "accuracy": 1.0,
+            "count": 108
+          }
+        }
+      },
+      "sub_M3": {
+        "full_accuracy": 1.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 179
+          },
+          "MB": {
+            "accuracy": 1.0,
+            "count": 103
+          },
+          "ME": {
+            "accuracy": 1.0,
+            "count": 56
+          },
+          "UB": {
+            "accuracy": 1.0,
+            "count": 149
+          },
+          "UD": {
+            "accuracy": 1.0,
+            "count": 213
+          }
+        }
+      },
+      "sub_M4": {
+        "full_accuracy": 1.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 200
+          },
+          "MB": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "UB": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "UD": {
+            "accuracy": 1.0,
+            "count": 300
+          }
+        }
+      },
+      "sub_M5": {
+        "full_accuracy": 0.14,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "MB": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "UB": {
+            "accuracy": 0.14,
+            "count": 100
+          },
+          "UD": {
+            "accuracy": 1.0,
+            "count": 400
+          }
+        }
+      },
+      "sub_random": {
+        "full_accuracy": 1.0,
+        "n_examples": 200,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 600
+          },
+          "MB": {
+            "accuracy": 1.0,
+            "count": 267
+          },
+          "ME": {
+            "accuracy": 1.0,
+            "count": 53
+          },
+          "UB": {
+            "accuracy": 1.0,
+            "count": 439
+          },
+          "UD": {
+            "accuracy": 1.0,
+            "count": 41
+          }
+        }
+      },
+      "sub_B3": {
+        "full_accuracy": 1.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 300
+          },
+          "MB": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "UB": {
+            "accuracy": 1.0,
+            "count": 197
+          },
+          "UD": {
+            "accuracy": 1.0,
+            "count": 103
+          }
+        }
+      },
+      "sub_B4": {
+        "full_accuracy": 1.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 200
+          },
+          "MB": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "UB": {
+            "accuracy": 1.0,
+            "count": 247
+          },
+          "UD": {
+            "accuracy": 1.0,
+            "count": 153
+          }
+        }
+      },
+      "sub_B5": {
+        "full_accuracy": 0.94,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "MB": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "UB": {
+            "accuracy": 0.9798657718120806,
+            "count": 298
+          },
+          "UD": {
+            "accuracy": 1.0,
+            "count": 202
+          }
+        }
+      }
+    },
+    "summary": {
+      "overall_accuracy": 0.96125,
+      "total_examples": 2400,
+      "n_splits": 22
+    }
+  },
+  "sorl_overall_accuracy": 0.96125,
+  "sft_overall_accuracy": 0.7683333333333333
+}

add_sub_sorl_v1_abs30_K1_10K/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d36997c56cc0957d831d187e34abff36823520d50c5fbaaa3cdca415484c9628
+size 650385300

add_sub_sorl_v1_abs30_K1_10K/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": 8e-05,
+  "emb_lr_mult": 1.0,
+  "weight_decay": 0.01,
+  "warmup_steps": 100,
+  "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": 20,
+  "emb_warmup_steps": 0,
+  "log_every": 50,
+  "eval_every": 156,
+  "save_every": 999999,
+  "eval_samples": 100,
+  "output_dir": "ckpt/sweep/as_sorl_abs30_K1_10K",
+  "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": 2,
+  "n_head": 3,
+  "n_embd": 510,
+  "ops": "add_sub",
+  "abs_vocab": 30,
+  "dataset_size": 10000,
+  "mode": "sorl",
+  "device": "cuda",
+  "push_to_hub": true,
+  "no_wandb": false,
+  "n_params": 162519662,
+  "run_name": "add_sub_sorl_v1_abs30_K1_10K",
+  "git_commit": "57deaa28d9c21e39ddac5ef448d6e1be992fba91",
+  "timestamp": "2026-04-13T09:57:42.064112+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": "6mbrz6zi",
+  "wandb_url": "https://wandb.ai/nlp_and_interpretability/sorl-arithmetic/runs/6mbrz6zi",
+  "final_accuracy": 0.96125,
+  "sft_accuracy": 0.7683333333333333,
+  "eval_method": "ArithmeticEvaluator"
+}