Upload add_sub_sorl_v1_abs20_K1_10K

add_sub_sorl_v1_abs20_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": 151664
+}

add_sub_sorl_v1_abs20_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_abs20_K1_10K/metrics.json

@@ -0,0 +1,1687 @@
+{
+  "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": [
+      6.609847068786621,
+      4.313945770263672,
+      2.8211135864257812,
+      2.6098971366882324,
+      2.8803181648254395,
+      2.4080252647399902,
+      2.382941722869873,
+      0.3024674654006958,
+      -4.526322364807129,
+      -6.631197452545166,
+      -4.271163463592529,
+      -2.845456838607788,
+      -1.7134591341018677,
+      -1.919076681137085,
+      -1.5664957761764526,
+      -1.3899976015090942,
+      -0.7998488545417786,
+      -1.1952388286590576,
+      -0.8307008743286133,
+      -0.5860740542411804,
+      -0.6104287505149841,
+      -0.8407340049743652,
+      -1.2119009494781494,
+      -1.0129437446594238,
+      -1.313117504119873,
+      -1.1106709241867065,
+      -1.6603138446807861,
+      -1.6090965270996094,
+      -1.0733165740966797,
+      -1.0066807270050049,
+      -1.2794181108474731,
+      -0.9843854904174805,
+      -1.123490333557129,
+      -1.0440940856933594,
+      -0.9606919884681702,
+      -1.414717674255371,
+      -0.8292436599731445,
+      -1.2538599967956543,
+      -0.919226884841919,
+      -0.5217360854148865,
+      -1.245757818222046,
+      -0.7945197820663452,
+      -0.6292770504951477,
+      -0.7419818639755249,
+      -0.5777878761291504,
+      -0.28315025568008423,
+      -0.35084664821624756,
+      -0.6482582092285156,
+      -0.33098843693733215,
+      -0.37686067819595337,
+      -0.5184847116470337,
+      -0.0973658561706543,
+      -0.3080046772956848,
+      -0.2875412106513977,
+      -0.20791345834732056,
+      -0.2233448624610901,
+      -0.16375748813152313,
+      -0.13954374194145203,
+      -0.12580332159996033,
+      -0.0063927918672561646
+    ],
+    "base_loss": [
+      7.595933437347412,
+      3.8396108150482178,
+      1.9869134426116943,
+      1.833506464958191,
+      1.7767173051834106,
+      1.8865851163864136,
+      1.8839255571365356,
+      1.7337242364883423,
+      1.8361998796463013,
+      1.6329604387283325,
+      1.2356112003326416,
+      0.9798861742019653,
+      0.7108522653579712,
+      0.6530479788780212,
+      0.5507955551147461,
+      0.5035597085952759,
+      0.3905493915081024,
+      0.356679767370224,
+      0.34553012251853943,
+      0.28289204835891724,
+      0.33556804060935974,
+      0.24587801098823547,
+      0.26413649320602417,
+      0.21149785816669464,
+      0.22125734388828278,
+      0.21185027062892914,
+      0.24059262871742249,
+      0.2523518204689026,
+      0.16996660828590393,
+      0.17189233005046844,
+      0.1756567806005478,
+      0.15790970623493195,
+      0.1541072428226471,
+      0.14071570336818695,
+      0.14612773060798645,
+      0.17851118743419647,
+      0.11484314501285553,
+      0.16218499839305878,
+      0.12426789849996567,
+      0.08009054511785507,
+      0.15919354557991028,
+      0.10112188756465912,
+      0.08774156123399734,
+      0.09424220025539398,
+      0.07499466091394424,
+      0.04172312468290329,
+      0.05625886842608452,
+      0.08554046601057053,
+      0.04787531867623329,
+      0.05242893472313881,
+      0.06812094151973724,
+      0.016909128054976463,
+      0.041645247489213943,
+      0.037941548973321915,
+      0.030796796083450317,
+      0.029599543660879135,
+      0.023928115144371986,
+      0.02092134952545166,
+      0.019392890855669975,
+      0.006476817186921835
+    ],
+    "info_loss": [
+      -0.6565923690795898,
+      -0.12502765655517578,
+      -0.049420952796936035,
+      -0.05284583568572998,
+      -0.02037215232849121,
+      -0.07843160629272461,
+      -0.08011436462402344,
+      -0.2734527587890625,
+      -0.7668462991714478,
+      -0.9567966461181641,
+      -0.6803109645843506,
+      -0.5087734460830688,
+      -0.3593854606151581,
+      -0.3683965802192688,
+      -0.3137052655220032,
+      -0.2846810519695282,
+      -0.21488648653030396,
+      -0.23822584748268127,
+      -0.18467622995376587,
+      -0.15311263501644135,
+      -0.1545877903699875,
+      -0.16041716933250427,
+      -0.1945183277130127,
+      -0.16658000648021698,
+      -0.18946321308612823,
+      -0.1700718104839325,
+      -0.22037822008132935,
+      -0.2142937034368515,
+      -0.15489357709884644,
+      -0.14260269701480865,
+      -0.16857698559761047,
+      -0.13929885625839233,
+      -0.14750534296035767,
+      -0.13791777193546295,
+      -0.13023759424686432,
+      -0.17602205276489258,
+      -0.11079932749271393,
+      -0.1599888950586319,
+      -0.11851393431425095,
+      -0.07508832216262817,
+      -0.15482710301876068,
+      -0.09955955296754837,
+      -0.08393968641757965,
+      -0.09361077845096588,
+      -0.07459273934364319,
+      -0.04118003323674202,
+      -0.04896640405058861,
+      -0.08194470405578613,
+      -0.04743151739239693,
+      -0.05198490992188454,
+      -0.06646489351987839,
+      -0.016638053581118584,
+      -0.04144047945737839,
+      -0.03774741664528847,
+      -0.030647873878479004,
+      -0.029441749677062035,
+      -0.02379968948662281,
+      -0.020802199840545654,
+      -0.019249044358730316,
+      -0.006381101440638304
+    ],
+    "abs_loss": [
+      2.9256432056427,
+      2.4958276748657227,
+      2.46895432472229,
+      2.4088761806488037,
+      2.4180774688720703,
+      2.4649078845977783,
+      2.4588210582733154,
+      2.453974723815918,
+      2.4324522018432617,
+      2.415642023086548,
+      2.352257490158081,
+      2.2088191509246826,
+      1.8896418809890747,
+      1.6739274263381958,
+      1.32405686378479,
+      1.0811830759048462,
+      1.0548852682113647,
+      0.758519172668457,
+      0.718182384967804,
+      0.6602980494499207,
+      0.6154454946517944,
+      0.46965107321739197,
+      0.44081011414527893,
+      0.4427615702152252,
+      0.35650837421417236,
+      0.3500964641571045,
+      0.3001129925251007,
+      0.2999572455883026,
+      0.25184276700019836,
+      0.2689480483531952,
+      0.2463725060224533,
+      0.24224205315113068,
+      0.2413814663887024,
+      0.20568279922008514,
+      0.1656782478094101,
+      0.19907349348068237,
+      0.16802407801151276,
+      0.17790837585926056,
+      0.16314257681369781,
+      0.17447948455810547,
+      0.1573050618171692,
+      0.13456518948078156,
+      0.12536819279193878,
+      0.10822665691375732,
+      0.09974483400583267,
+      0.11742156744003296,
+      0.11229091882705688,
+      0.09441050887107849,
+      0.08244535326957703,
+      0.08592605590820312,
+      0.07963317632675171,
+      0.08054837584495544,
+      0.07055239379405975,
+      0.089305579662323,
+      0.07961511611938477,
+      0.08294523507356644,
+      0.0843416154384613,
+      0.07325228303670883,
+      0.07107876986265182,
+      0.08012231439352036
+    ],
+    "zipf_loss": [
+      5.287272930145264,
+      1.4750287532806396,
+      1.0815143585205078,
+      1.063961386680603,
+      1.0655146837234497,
+      1.059265375137329,
+      1.0542776584625244,
+      1.0578733682632446,
+      1.0626957416534424,
+      1.0622444152832031,
+      1.0611090660095215,
+      1.0415093898773193,
+      0.9805790185928345,
+      0.9444482326507568,
+      0.8873554468154907,
+      0.8451348543167114,
+      0.8529779314994812,
+      0.7544879913330078,
+      0.5987130403518677,
+      0.5961304306983948,
+      0.5383365750312805,
+      0.4705946445465088,
+      0.4250647723674774,
+      0.3970823884010315,
+      0.32460635900497437,
+      0.3431873321533203,
+      0.27286437153816223,
+      0.2514929175376892,
+      0.2804683446884155,
+      0.22055907547473907,
+      0.20605769753456116,
+      0.22646912932395935,
+      0.17331784963607788,
+      0.17379960417747498,
+      0.17898838222026825,
+      0.14708437025547028,
+      0.14710399508476257,
+      0.16605313122272491,
+      0.12533023953437805,
+      0.13160867989063263,
+      0.12758919596672058,
+      0.08649729937314987,
+      0.10984141379594803,
+      0.08906104415655136,
+      0.08317035436630249,
+      0.07518477737903595,
+      0.07132942974567413,
+      0.07620728760957718,
+      0.0872068703174591,
+      0.08196690678596497,
+      0.07007996737957001,
+      0.04405072703957558,
+      0.05769962817430496,
+      0.043060846626758575,
+      0.05980696901679039,
+      0.03317856416106224,
+      0.04187712073326111,
+      0.040231671184301376,
+      0.04018634557723999,
+      0.04292917251586914
+    ],
+    "denoise_loss": [],
+    "ortho_loss": [
+      0.35471978783607483,
+      0.10054542869329453,
+      0.08056585490703583,
+      0.06761873513460159,
+      0.06051715835928917,
+      0.054626185446977615,
+      0.06290311366319656,
+      0.08528075367212296,
+      0.12851166725158691,
+      0.17594954371452332,
+      0.1883549988269806,
+      0.20819245278835297,
+      0.22447381913661957,
+      0.23362860083580017,
+      0.23352178931236267,
+      0.2505631148815155,
+      0.2545468807220459,
+      0.25482049584388733,
+      0.2553919851779938,
+      0.2531490623950958,
+      0.2515760660171509,
+      0.25270405411720276,
+      0.2550104260444641,
+      0.24405382573604584,
+      0.2355482131242752,
+      0.23306143283843994,
+      0.2306639552116394,
+      0.23312798142433167,
+      0.2199832797050476,
+      0.2244483232498169,
+      0.22459453344345093,
+      0.2254096269607544,
+      0.22532722353935242,
+      0.22544094920158386,
+      0.22568713128566742,
+      0.22397875785827637,
+      0.232473224401474,
+      0.23453326523303986,
+      0.2402927279472351,
+      0.24208620190620422,
+      0.23906201124191284,
+      0.2491123527288437,
+      0.2513620853424072,
+      0.24948538839817047,
+      0.25132355093955994,
+      0.2519170641899109,
+      0.25072360038757324,
+      0.25337281823158264,
+      0.25460267066955566,
+      0.2526955306529999,
+      0.2516002357006073,
+      0.252238005399704,
+      0.2524856626987457,
+      0.2535747289657593,
+      0.256831556558609,
+      0.2574509084224701,
+      0.2587626278400421,
+      0.25879767537117004,
+      0.25963112711906433,
+      0.2598315477371216
+    ],
+    "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": [
+      0.02,
+      0.01,
+      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
+    ]
+  },
+  "final_accuracy": 0.9519230769230769,
+  "sft_eval": {
+    "config": {
+      "ops": "add_sub",
+      "K": null,
+      "mode": "sft",
+      "n_digits": 6,
+      "n_per_split": 100
+    },
+    "splits": {
+      "add_S0": {
+        "full_accuracy": 1.0,
+        "digit_accuracy": 1.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 605
+          },
+          "SS": {
+            "accuracy": 1.0,
+            "count": 95
+          }
+        }
+      },
+      "add_S1": {
+        "full_accuracy": 0.98,
+        "digit_accuracy": 0.9971428571428571,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 204
+          },
+          "SC": {
+            "accuracy": 0.9940828402366864,
+            "count": 169
+          },
+          "SS": {
+            "accuracy": 1.0,
+            "count": 31
+          },
+          "UC": {
+            "accuracy": 0.9966216216216216,
+            "count": 296
+          }
+        }
+      },
+      "add_S2": {
+        "full_accuracy": 0.99,
+        "digit_accuracy": 0.9985714285714286,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 163
+          },
+          "SC": {
+            "accuracy": 1.0,
+            "count": 130
+          },
+          "SS": {
+            "accuracy": 0.9885057471264368,
+            "count": 87
+          },
+          "UC": {
+            "accuracy": 1.0,
+            "count": 203
+          },
+          "US": {
+            "accuracy": 1.0,
+            "count": 117
+          }
+        }
+      },
+      "add_S3": {
+        "full_accuracy": 0.83,
+        "digit_accuracy": 0.9757142857142858,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 121
+          },
+          "SC": {
+            "accuracy": 0.9917355371900827,
+            "count": 121
+          },
+          "SS": {
+            "accuracy": 1.0,
+            "count": 49
+          },
+          "UC": {
+            "accuracy": 0.9139784946236559,
+            "count": 186
+          },
+          "US": {
+            "accuracy": 1.0,
+            "count": 223
+          }
+        }
+      },
+      "add_S4": {
+        "full_accuracy": 0.77,
+        "digit_accuracy": 0.9585714285714285,
+        "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": 0.8875,
+            "count": 160
+          },
+          "US": {
+            "accuracy": 0.9641693811074918,
+            "count": 307
+          }
+        }
+      },
+      "add_S5": {
+        "full_accuracy": 0.54,
+        "digit_accuracy": 0.8628571428571429,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "SC": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "UC": {
+            "accuracy": 0.69,
+            "count": 100
+          },
+          "US": {
+            "accuracy": 0.8375,
+            "count": 400
+          }
+        }
+      },
+      "add_S6": {
+        "full_accuracy": 0.91,
+        "digit_accuracy": 0.9657142857142857,
+        "n_examples": 100,
+        "per_subtask": {
+          "SC": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "UC": {
+            "accuracy": 0.96,
+            "count": 100
+          },
+          "US": {
+            "accuracy": 0.96,
+            "count": 500
+          }
+        }
+      },
+      "add_random": {
+        "full_accuracy": 0.985,
+        "digit_accuracy": 0.9978571428571429,
+        "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": 0.994328922495274,
+            "count": 529
+          },
+          "US": {
+            "accuracy": 1.0,
+            "count": 48
+          }
+        }
+      },
+      "add_C1": {
+        "full_accuracy": 1.0,
+        "digit_accuracy": 1.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 500
+          },
+          "SC": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "UC": {
+            "accuracy": 1.0,
+            "count": 100
+          }
+        }
+      },
+      "add_C2": {
+        "full_accuracy": 0.98,
+        "digit_accuracy": 0.9971428571428571,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 0.9975,
+            "count": 400
+          },
+          "SC": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "UC": {
+            "accuracy": 0.9935897435897436,
+            "count": 156
+          },
+          "US": {
+            "accuracy": 1.0,
+            "count": 44
+          }
+        }
+      },
+      "add_C3": {
+        "full_accuracy": 0.9,
+        "digit_accuracy": 0.9857142857142858,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 300
+          },
+          "SC": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "UC": {
+            "accuracy": 0.949748743718593,
+            "count": 199
+          },
+          "US": {
+            "accuracy": 1.0,
+            "count": 101
+          }
+        }
+      },
+      "add_C4": {
+        "full_accuracy": 0.88,
+        "digit_accuracy": 0.9828571428571429,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 200
+          },
+          "SC": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "UC": {
+            "accuracy": 0.9545454545454546,
+            "count": 264
+          },
+          "US": {
+            "accuracy": 1.0,
+            "count": 136
+          }
+        }
+      },
+      "add_C5": {
+        "full_accuracy": 0.83,
+        "digit_accuracy": 0.97,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "SC": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "UC": {
+            "accuracy": 0.9483870967741935,
+            "count": 310
+          },
+          "US": {
+            "accuracy": 0.9736842105263158,
+            "count": 190
+          }
+        }
+      },
+      "add_C6": {
+        "full_accuracy": 0.84,
+        "digit_accuracy": 0.9685714285714285,
+        "n_examples": 100,
+        "per_subtask": {
+          "SC": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "UC": {
+            "accuracy": 0.9675675675675676,
+            "count": 370
+          },
+          "US": {
+            "accuracy": 0.9565217391304348,
+            "count": 230
+          }
+        }
+      },
+      "sub_M0": {
+        "full_accuracy": 1.0,
+        "digit_accuracy": 1.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 615
+          },
+          "ME": {
+            "accuracy": 1.0,
+            "count": 85
+          }
+        }
+      },
+      "sub_M1": {
+        "full_accuracy": 1.0,
+        "digit_accuracy": 1.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 292
+          },
+          "MB": {
+            "accuracy": 1.0,
+            "count": 144
+          },
+          "ME": {
+            "accuracy": 1.0,
+            "count": 25
+          },
+          "UB": {
+            "accuracy": 1.0,
+            "count": 239
+          }
+        }
+      },
+      "sub_M2": {
+        "full_accuracy": 1.0,
+        "digit_accuracy": 1.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 211
+          },
+          "MB": {
+            "accuracy": 1.0,
+            "count": 115
+          },
+          "ME": {
+            "accuracy": 1.0,
+            "count": 85
+          },
+          "UB": {
+            "accuracy": 1.0,
+            "count": 181
+          },
+          "UD": {
+            "accuracy": 1.0,
+            "count": 108
+          }
+        }
+      },
+      "sub_M3": {
+        "full_accuracy": 0.84,
+        "digit_accuracy": 0.9771428571428571,
+        "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": 0.8926174496644296,
+            "count": 149
+          },
+          "UD": {
+            "accuracy": 1.0,
+            "count": 213
+          }
+        }
+      },
+      "sub_M4": {
+        "full_accuracy": 0.13,
+        "digit_accuracy": 0.81,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 200
+          },
+          "MB": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "UB": {
+            "accuracy": 0.14,
+            "count": 100
+          },
+          "UD": {
+            "accuracy": 0.8433333333333334,
+            "count": 300
+          }
+        }
+      },
+      "sub_M5": {
+        "full_accuracy": 0.08,
+        "digit_accuracy": 0.7042857142857143,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "MB": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "UB": {
+            "accuracy": 0.27,
+            "count": 100
+          },
+          "UD": {
+            "accuracy": 0.665,
+            "count": 400
+          }
+        }
+      },
+      "sub_random": {
+        "full_accuracy": 1.0,
+        "digit_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": 0.84,
+        "digit_accuracy": 0.9771428571428571,
+        "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": 0.79,
+        "digit_accuracy": 0.9642857142857143,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 200
+          },
+          "MB": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "UB": {
+            "accuracy": 0.9149797570850202,
+            "count": 247
+          },
+          "UD": {
+            "accuracy": 0.9738562091503268,
+            "count": 153
+          }
+        }
+      },
+      "sub_B5": {
+        "full_accuracy": 0.84,
+        "digit_accuracy": 0.9614285714285714,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "MB": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "UB": {
+            "accuracy": 0.9496644295302014,
+            "count": 298
+          },
+          "UD": {
+            "accuracy": 0.9405940594059405,
+            "count": 202
+          }
+        }
+      }
+    },
+    "summary": {
+      "overall_accuracy": 0.8438461538461538,
+      "digit_accuracy": 0.9634615384615385,
+      "total_examples": 2600,
+      "n_splits": 24
+    }
+  },
+  "sorl_eval": {
+    "config": {
+      "ops": "add_sub",
+      "K": 1,
+      "mode": "sorl",
+      "n_digits": 6,
+      "n_per_split": 100
+    },
+    "splits": {
+      "add_S0": {
+        "full_accuracy": 1.0,
+        "digit_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,
+        "digit_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,
+        "digit_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,
+        "digit_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,
+        "digit_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": 0.56,
+        "digit_accuracy": 0.9357142857142857,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "SC": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "UC": {
+            "accuracy": 0.56,
+            "count": 100
+          },
+          "US": {
+            "accuracy": 0.9975,
+            "count": 400
+          }
+        }
+      },
+      "add_S6": {
+        "full_accuracy": 1.0,
+        "digit_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,
+        "digit_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_C1": {
+        "full_accuracy": 1.0,
+        "digit_accuracy": 1.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 500
+          },
+          "SC": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "UC": {
+            "accuracy": 1.0,
+            "count": 100
+          }
+        }
+      },
+      "add_C2": {
+        "full_accuracy": 1.0,
+        "digit_accuracy": 1.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 400
+          },
+          "SC": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "UC": {
+            "accuracy": 1.0,
+            "count": 156
+          },
+          "US": {
+            "accuracy": 1.0,
+            "count": 44
+          }
+        }
+      },
+      "add_C3": {
+        "full_accuracy": 1.0,
+        "digit_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": 199
+          },
+          "US": {
+            "accuracy": 1.0,
+            "count": 101
+          }
+        }
+      },
+      "add_C4": {
+        "full_accuracy": 0.99,
+        "digit_accuracy": 0.9985714285714286,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 200
+          },
+          "SC": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "UC": {
+            "accuracy": 0.9962121212121212,
+            "count": 264
+          },
+          "US": {
+            "accuracy": 1.0,
+            "count": 136
+          }
+        }
+      },
+      "add_C5": {
+        "full_accuracy": 0.99,
+        "digit_accuracy": 0.9985714285714286,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "SC": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "UC": {
+            "accuracy": 0.9967741935483871,
+            "count": 310
+          },
+          "US": {
+            "accuracy": 1.0,
+            "count": 190
+          }
+        }
+      },
+      "add_C6": {
+        "full_accuracy": 0.96,
+        "digit_accuracy": 0.9942857142857143,
+        "n_examples": 100,
+        "per_subtask": {
+          "SC": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "UC": {
+            "accuracy": 0.9891891891891892,
+            "count": 370
+          },
+          "US": {
+            "accuracy": 1.0,
+            "count": 230
+          }
+        }
+      },
+      "sub_M0": {
+        "full_accuracy": 1.0,
+        "digit_accuracy": 1.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 615
+          },
+          "ME": {
+            "accuracy": 1.0,
+            "count": 85
+          }
+        }
+      },
+      "sub_M1": {
+        "full_accuracy": 1.0,
+        "digit_accuracy": 1.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 292
+          },
+          "MB": {
+            "accuracy": 1.0,
+            "count": 144
+          },
+          "ME": {
+            "accuracy": 1.0,
+            "count": 25
+          },
+          "UB": {
+            "accuracy": 1.0,
+            "count": 239
+          }
+        }
+      },
+      "sub_M2": {
+        "full_accuracy": 1.0,
+        "digit_accuracy": 1.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 211
+          },
+          "MB": {
+            "accuracy": 1.0,
+            "count": 115
+          },
+          "ME": {
+            "accuracy": 1.0,
+            "count": 85
+          },
+          "UB": {
+            "accuracy": 1.0,
+            "count": 181
+          },
+          "UD": {
+            "accuracy": 1.0,
+            "count": 108
+          }
+        }
+      },
+      "sub_M3": {
+        "full_accuracy": 1.0,
+        "digit_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": 0.98,
+        "digit_accuracy": 0.9971428571428571,
+        "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": 0.9933333333333333,
+            "count": 300
+          }
+        }
+      },
+      "sub_M5": {
+        "full_accuracy": 0.34,
+        "digit_accuracy": 0.9057142857142857,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "MB": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "UB": {
+            "accuracy": 0.34,
+            "count": 100
+          },
+          "UD": {
+            "accuracy": 1.0,
+            "count": 400
+          }
+        }
+      },
+      "sub_random": {
+        "full_accuracy": 1.0,
+        "digit_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": 0.99,
+        "digit_accuracy": 0.9985714285714286,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 300
+          },
+          "MB": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "UB": {
+            "accuracy": 0.9949238578680203,
+            "count": 197
+          },
+          "UD": {
+            "accuracy": 1.0,
+            "count": 103
+          }
+        }
+      },
+      "sub_B4": {
+        "full_accuracy": 1.0,
+        "digit_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,
+        "digit_accuracy": 0.9914285714285714,
+        "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": 0.9519230769230769,
+      "digit_accuracy": 0.9930769230769231,
+      "total_examples": 2600,
+      "n_splits": 24
+    }
+  },
+  "sorl_overall_accuracy": 0.9519230769230769,
+  "sft_overall_accuracy": 0.8438461538461538
+}

add_sub_sorl_v1_abs20_K1_10K/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:22e372b9ada964d440d4df2367917a51ae15eb6d718813ee947798a1086f11d6
+size 650344480

add_sub_sorl_v1_abs20_K1_10K/train_config.json

@@ -0,0 +1,89 @@
+{
+  "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_abs20_K1_10K_2L3H510d",
+  "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": 20,
+  "dataset_size": 10000,
+  "mode": "sorl",
+  "device": "cuda",
+  "push_to_hub": true,
+  "no_wandb": false,
+  "n_params": 162509462,
+  "run_name": "add_sub_sorl_v1_abs20_K1_10K",
+  "git_commit": "f835493c19eb98267697007042c9d440cad2afbb",
+  "timestamp": "2026-04-16T04:10:46.561234+00:00",
+  "tokenizer": "Qwen/Qwen3-0.6B",
+  "dataset_repo": "thoughtworks/arithmetic-sorl-data",
+  "dataset_config": "add_sub_6digit",
+  "train_dataset": "fixed_train/train_10K_seed42.pt",
+  "model_repo": "thoughtworks/arithmetic-sorl",
+  "trainer_version": "v1",
+  "wandb_run_id": "o9oa527y",
+  "wandb_url": "https://wandb.ai/nlp_and_interpretability/sorl-arithmetic/runs/o9oa527y",
+  "eval_final_dataset": "eval_sets/eval_add_sub_6d_N100_seed42.json",
+  "eval_epoch_dataset": "eval_sets/eval_add_sub_6d_N25_seed42.json",
+  "eval_hf_repo": "thoughtworks/arithmetic-sorl-data",
+  "config_hash": "8952e8187a2b",
+  "final_accuracy": 0.9519230769230769,
+  "sft_accuracy": 0.8438461538461538,
+  "eval_method": "ArithmeticEvaluator"
+}