Upload add_sub_sorl_v1_abs2_K1_10K

add_sub_sorl_v1_abs2_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": 151646
+}

add_sub_sorl_v1_abs2_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_abs2_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": [
+      12.146018028259277,
+      8.140860557556152,
+      5.788841247558594,
+      5.928689956665039,
+      4.940187931060791,
+      -0.8538930416107178,
+      -3.957218647003174,
+      -5.3563971519470215,
+      -6.507109642028809,
+      -7.65699577331543,
+      -8.334919929504395,
+      -9.785987854003906,
+      -9.972005844116211,
+      -11.117347717285156,
+      -11.665057182312012,
+      -10.518767356872559,
+      -12.363883018493652,
+      -11.287307739257812,
+      -9.55109977722168,
+      -7.357775688171387,
+      -5.654560565948486,
+      -3.8458449840545654,
+      -1.607912540435791,
+      -1.7796285152435303,
+      -1.6869027614593506,
+      -1.2786306142807007,
+      -1.1483945846557617,
+      -0.8021516799926758,
+      -0.6225531101226807,
+      -0.4951012432575226,
+      -1.253838062286377,
+      -0.6078148484230042,
+      -0.5363032817840576,
+      -0.4771603047847748,
+      -0.4060630202293396,
+      -0.5651876926422119,
+      -0.4811541736125946,
+      -0.5336906313896179,
+      -0.29910436272621155,
+      -0.1743965446949005,
+      -0.0865909680724144,
+      -0.19921013712882996,
+      -0.09337317198514938,
+      -0.33079972863197327,
+      -0.22505944967269897,
+      -0.13146325945854187,
+      -0.16447202861309052,
+      -0.11612369120121002,
+      0.009087484329938889,
+      0.03794008493423462,
+      0.02065785601735115,
+      0.0008186213672161102,
+      -0.01969563215970993,
+      0.0013124383985996246,
+      0.02527482435107231,
+      0.05165558308362961,
+      0.051053162664175034,
+      0.04289375990629196,
+      0.05697126314043999,
+      0.05128497630357742
+    ],
+    "base_loss": [
+      7.628437042236328,
+      4.0556206703186035,
+      1.9530707597732544,
+      1.8979682922363281,
+      1.9241129159927368,
+      1.8852487802505493,
+      1.8353227376937866,
+      1.8186378479003906,
+      1.8541980981826782,
+      1.817212462425232,
+      1.7589912414550781,
+      1.752821683883667,
+      1.7150675058364868,
+      1.7124515771865845,
+      1.7245622873306274,
+      1.5408344268798828,
+      1.5466477870941162,
+      1.374353051185608,
+      1.1373178958892822,
+      0.8995004296302795,
+      0.6910373568534851,
+      0.4638596177101135,
+      0.231327623128891,
+      0.24538195133209229,
+      0.20177486538887024,
+      0.15957574546337128,
+      0.15221306681632996,
+      0.10463942587375641,
+      0.10045579820871353,
+      0.06593646109104156,
+      0.15611542761325836,
+      0.07823602110147476,
+      0.08280634135007858,
+      0.06120121851563454,
+      0.05883612856268883,
+      0.07205647975206375,
+      0.0748705118894577,
+      0.06757231056690216,
+      0.04013211280107498,
+      0.02632850408554077,
+      0.016306018456816673,
+      0.029043469578027725,
+      0.016971977427601814,
+      0.04336889460682869,
+      0.03189310058951378,
+      0.0211532860994339,
+      0.02499561943113804,
+      0.019456015899777412,
+      0.005540237762033939,
+      0.0022714717779308558,
+      0.004181382246315479,
+      0.006391526199877262,
+      0.00866654235869646,
+      0.006326296832412481,
+      0.0036668891552835703,
+      0.0007290642824955285,
+      0.0007950459257699549,
+      0.0017015466000884771,
+      0.00013872093404643238,
+      0.000770941493101418
+    ],
+    "info_loss": [
+      -0.6176466941833496,
+      -0.08461570739746094,
+      -0.05701446533203125,
+      -0.035248637199401855,
+      -0.13141179084777832,
+      -0.6646270751953125,
+      -0.9369587898254395,
+      -1.0392465591430664,
+      -1.134619951248169,
+      -1.2192878723144531,
+      -1.2667748928070068,
+      -1.4063975811004639,
+      -1.4086241722106934,
+      -1.5143393278121948,
+      -1.5569674968719482,
+      -1.3761557340621948,
+      -1.4464160203933716,
+      -1.2794716358184814,
+      -1.0764862298965454,
+      -0.8341529369354248,
+      -0.6416376829147339,
+      -0.4403914213180542,
+      -0.1967826783657074,
+      -0.2110116183757782,
+      -0.19503700733184814,
+      -0.14979569613933563,
+      -0.1361372172832489,
+      -0.09671700745820999,
+      -0.07853375375270844,
+      -0.06197825074195862,
+      -0.14958061277866364,
+      -0.07456944137811661,
+      -0.0677691251039505,
+      -0.059750691056251526,
+      -0.05245509371161461,
+      -0.06968473643064499,
+      -0.06143011525273323,
+      -0.06598454713821411,
+      -0.03974508121609688,
+      -0.02589520253241062,
+      -0.016123393550515175,
+      -0.028640657663345337,
+      -0.01684783399105072,
+      -0.043257907032966614,
+      -0.03150363266468048,
+      -0.021070921793580055,
+      -0.02475244365632534,
+      -0.019364137202501297,
+      -0.005449273623526096,
+      -0.002236021449789405,
+      -0.004155138973146677,
+      -0.006360492669045925,
+      -0.008639117702841759,
+      -0.0063036042265594006,
+      -0.0036412517074495554,
+      -0.0007088789134286344,
+      -0.000775613822042942,
+      -0.0016821041936054826,
+      -0.00011882289254572242,
+      -0.0007504023960791528
+    ],
+    "abs_loss": [
+      0.5970545411109924,
+      0.07690786570310593,
+      0.0035618257243186235,
+      0.0012020958820357919,
+      0.001016044756397605,
+      0.002263190457597375,
+      0.008593530394136906,
+      0.012380805797874928,
+      0.01136107463389635,
+      0.021077027544379234,
+      0.015056786127388477,
+      0.015377345494925976,
+      0.01237007137387991,
+      0.009393605403602123,
+      0.010074876248836517,
+      0.008252273313701153,
+      0.010502271354198456,
+      0.005695981439203024,
+      0.006076695863157511,
+      0.004774859175086021,
+      0.003744877176359296,
+      0.002134924056008458,
+      0.002635170938447118,
+      0.001622178009711206,
+      0.0014428232097998261,
+      0.001420547254383564,
+      0.0018245340324938297,
+      0.0009455258841626346,
+      0.001069836551323533,
+      0.0005867307772859931,
+      0.0007147423457354307,
+      0.00043318382813595235,
+      0.00048265376244671643,
+      0.0004175094945821911,
+      0.00023684576444793493,
+      0.00024159869644790888,
+      0.00029623997397720814,
+      0.0002410489832982421,
+      0.0002485735749360174,
+      0.00016171969764400274,
+      9.246024274034426e-05,
+      0.00010874952567974105,
+      9.772043995326385e-05,
+      0.00010663767898222432,
+      9.5514660642948e-05,
+      6.556689186254516e-05,
+      9.23834159038961e-05,
+      5.149324351805262e-05,
+      5.457560837385245e-05,
+      3.461950836936012e-05,
+      3.3307816920569167e-05,
+      3.955264764954336e-05,
+      3.886878039338626e-05,
+      3.2628566259518266e-05,
+      2.6662886739359237e-05,
+      2.7881047572009265e-05,
+      1.9944345694966614e-05,
+      2.5962357540265657e-05,
+      2.4488110284437425e-05,
+      2.169463004975114e-05
+    ],
+    "zipf_loss": [
+      10.634342193603516,
+      4.9237060546875,
+      4.405559062957764,
+      4.383087635040283,
+      4.33009147644043,
+      3.906902551651001,
+      3.5761871337890625,
+      3.2161927223205566,
+      2.983755111694336,
+      2.7165632247924805,
+      2.5723321437835693,
+      2.523629665374756,
+      2.3979315757751465,
+      2.3126540184020996,
+      2.1790482997894287,
+      1.7011308670043945,
+      0.5525786280632019,
+      0.1324852854013443,
+      0.07583678513765335,
+      0.08377665281295776,
+      0.07040485739707947,
+      0.09399650245904922,
+      0.12832312285900116,
+      0.084943488240242,
+      0.06154828518629074,
+      0.0596085824072361,
+      0.060582030564546585,
+      0.06028445437550545,
+      0.062221676111221313,
+      0.05868617817759514,
+      0.08578114211559296,
+      0.05960015952587128,
+      0.05853329226374626,
+      0.059103649109601974,
+      0.059628069400787354,
+      0.05957905203104019,
+      0.05824682116508484,
+      0.05855847895145416,
+      0.05818948894739151,
+      0.05821080133318901,
+      0.05832771211862564,
+      0.058142099529504776,
+      0.05812341719865799,
+      0.05839976668357849,
+      0.058074235916137695,
+      0.058086104691028595,
+      0.058047544211149216,
+      0.058056507259607315,
+      0.05803452432155609,
+      0.05802536383271217,
+      0.058024533092975616,
+      0.05802806839346886,
+      0.058025114238262177,
+      0.058018919080495834,
+      0.0580177865922451,
+      0.05801251903176308,
+      0.05801226198673248,
+      0.058010656386613846,
+      0.05801832303404808,
+      0.058015890419483185
+    ],
+    "denoise_loss": [],
+    "ortho_loss": [
+      0.12159737944602966,
+      0.18038727343082428,
+      0.03381244093179703,
+      0.0021874275989830494,
+      2.175434747186955e-05,
+      0.00040765784797258675,
+      0.0030973420944064856,
+      0.00593204190954566,
+      0.0047910804860293865,
+      0.005361207760870457,
+      3.893865709869715e-07,
+      0.0006737676449120045,
+      1.8460258388586226e-07,
+      0.0002703091886360198,
+      1.5977379916876089e-06,
+      0.0005756249302066863,
+      0.0026328135281801224,
+      0.0036702558863908052,
+      0.002528548240661621,
+      0.003325717756524682,
+      0.004215314518660307,
+      0.005500412080436945,
+      0.011995360255241394,
+      0.015324790962040424,
+      0.02708487957715988,
+      0.025101546198129654,
+      0.027320556342601776,
+      0.03245297074317932,
+      0.03193245083093643,
+      0.03052922524511814,
+      0.029022295027971268,
+      0.026948127895593643,
+      0.030607450753450394,
+      0.028174009174108505,
+      0.03190726414322853,
+      0.02512122131884098,
+      0.11264711618423462,
+      0.11880636215209961,
+      0.11841752380132675,
+      0.1180657371878624,
+      0.11860547959804535,
+      0.11567839980125427,
+      0.11602658033370972,
+      0.11348988115787506,
+      0.11054149270057678,
+      0.1121196448802948,
+      0.11147750914096832,
+      0.11120431870222092,
+      0.11274542659521103,
+      0.1115327998995781,
+      0.11288487911224365,
+      0.11170242726802826,
+      0.11522012948989868,
+      0.11410203576087952,
+      0.11274243891239166,
+      0.11297151446342468,
+      0.11337342858314514,
+      0.11387932300567627,
+      0.11410986632108688,
+      0.1143156886100769
+    ],
+    "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,
+      0.02,
+      0.02,
+      0.02,
+      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": 0.9919230769230769,
+  "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": 0.99,
+        "digit_accuracy": 0.9985714285714286,
+        "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": 0.9966216216216216,
+            "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": 0.99,
+        "digit_accuracy": 0.9985714285714286,
+        "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": 0.9946236559139785,
+            "count": 186
+          },
+          "US": {
+            "accuracy": 1.0,
+            "count": 223
+          }
+        }
+      },
+      "add_S4": {
+        "full_accuracy": 0.99,
+        "digit_accuracy": 0.9985714285714286,
+        "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": 0.99375,
+            "count": 160
+          },
+          "US": {
+            "accuracy": 1.0,
+            "count": 307
+          }
+        }
+      },
+      "add_S5": {
+        "full_accuracy": 0.85,
+        "digit_accuracy": 0.9785714285714285,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "SC": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "UC": {
+            "accuracy": 0.85,
+            "count": 100
+          },
+          "US": {
+            "accuracy": 1.0,
+            "count": 400
+          }
+        }
+      },
+      "add_S6": {
+        "full_accuracy": 0.95,
+        "digit_accuracy": 0.9928571428571429,
+        "n_examples": 100,
+        "per_subtask": {
+          "SC": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "UC": {
+            "accuracy": 0.95,
+            "count": 100
+          },
+          "US": {
+            "accuracy": 1.0,
+            "count": 500
+          }
+        }
+      },
+      "add_random": {
+        "full_accuracy": 0.995,
+        "digit_accuracy": 0.9992857142857143,
+        "n_examples": 200,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 447
+          },
+          "SC": {
+            "accuracy": 1.0,
+            "count": 320
+          },
+          "SS": {
+            "accuracy": 0.9821428571428571,
+            "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": 1.0,
+        "digit_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": 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": 1.0,
+        "digit_accuracy": 1.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "SC": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "UC": {
+            "accuracy": 1.0,
+            "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": 0.99,
+        "digit_accuracy": 0.9985714285714286,
+        "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": 0.99581589958159,
+            "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": 0.99,
+        "digit_accuracy": 0.9985714285714286,
+        "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": 0.9932885906040269,
+            "count": 149
+          },
+          "UD": {
+            "accuracy": 1.0,
+            "count": 213
+          }
+        }
+      },
+      "sub_M4": {
+        "full_accuracy": 0.45,
+        "digit_accuracy": 0.9214285714285714,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 200
+          },
+          "MB": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "UB": {
+            "accuracy": 0.45,
+            "count": 100
+          },
+          "UD": {
+            "accuracy": 1.0,
+            "count": 300
+          }
+        }
+      },
+      "sub_M5": {
+        "full_accuracy": 0.56,
+        "digit_accuracy": 0.92,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "MB": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "UB": {
+            "accuracy": 0.81,
+            "count": 100
+          },
+          "UD": {
+            "accuracy": 0.9075,
+            "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": 0.92,
+        "digit_accuracy": 0.9885714285714285,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 0.995,
+            "count": 200
+          },
+          "MB": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "UB": {
+            "accuracy": 0.97165991902834,
+            "count": 247
+          },
+          "UD": {
+            "accuracy": 1.0,
+            "count": 153
+          }
+        }
+      },
+      "sub_B5": {
+        "full_accuracy": 0.95,
+        "digit_accuracy": 0.99,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "MB": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "UB": {
+            "accuracy": 0.9899328859060402,
+            "count": 298
+          },
+          "UD": {
+            "accuracy": 0.9801980198019802,
+            "count": 202
+          }
+        }
+      }
+    },
+    "summary": {
+      "overall_accuracy": 0.9461538461538461,
+      "digit_accuracy": 0.9914835164835165,
+      "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": 0.99,
+        "digit_accuracy": 0.9985714285714286,
+        "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": 0.9946236559139785,
+            "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": 0.96,
+        "digit_accuracy": 0.9942857142857143,
+        "n_examples": 100,
+        "per_subtask": {
+          "SA": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "SC": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "UC": {
+            "accuracy": 0.96,
+            "count": 100
+          },
+          "US": {
+            "accuracy": 1.0,
+            "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": 1.0,
+        "digit_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": 264
+          },
+          "US": {
+            "accuracy": 1.0,
+            "count": 136
+          }
+        }
+      },
+      "add_C5": {
+        "full_accuracy": 1.0,
+        "digit_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": 310
+          },
+          "US": {
+            "accuracy": 1.0,
+            "count": 190
+          }
+        }
+      },
+      "add_C6": {
+        "full_accuracy": 1.0,
+        "digit_accuracy": 1.0,
+        "n_examples": 100,
+        "per_subtask": {
+          "SC": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "UC": {
+            "accuracy": 1.0,
+            "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": 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": 100
+          },
+          "UD": {
+            "accuracy": 1.0,
+            "count": 300
+          }
+        }
+      },
+      "sub_M5": {
+        "full_accuracy": 0.87,
+        "digit_accuracy": 0.9814285714285714,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "MB": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "UB": {
+            "accuracy": 0.87,
+            "count": 100
+          },
+          "UD": {
+            "accuracy": 1.0,
+            "count": 400
+          }
+        }
+      },
+      "sub_random": {
+        "full_accuracy": 0.995,
+        "digit_accuracy": 0.9992857142857143,
+        "n_examples": 200,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 0.9983333333333333,
+            "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,
+        "digit_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": 0.99,
+        "digit_accuracy": 0.9985714285714286,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 200
+          },
+          "MB": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "UB": {
+            "accuracy": 0.9959514170040485,
+            "count": 247
+          },
+          "UD": {
+            "accuracy": 1.0,
+            "count": 153
+          }
+        }
+      },
+      "sub_B5": {
+        "full_accuracy": 0.99,
+        "digit_accuracy": 0.9985714285714286,
+        "n_examples": 100,
+        "per_subtask": {
+          "MD": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "MB": {
+            "accuracy": 1.0,
+            "count": 100
+          },
+          "UB": {
+            "accuracy": 0.9966442953020134,
+            "count": 298
+          },
+          "UD": {
+            "accuracy": 1.0,
+            "count": 202
+          }
+        }
+      }
+    },
+    "summary": {
+      "overall_accuracy": 0.9919230769230769,
+      "digit_accuracy": 0.9988461538461538,
+      "total_examples": 2600,
+      "n_splits": 24
+    }
+  },
+  "sorl_overall_accuracy": 0.9919230769230769,
+  "sft_overall_accuracy": 0.9461538461538461
+}

add_sub_sorl_v1_abs2_K1_10K/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:556319ae59912bb52ef9f1eb1cf60b8fa2147c5b065a949f63161676310ad6ee
+size 650271004

add_sub_sorl_v1_abs2_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_abs2_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": 2,
+  "dataset_size": 10000,
+  "mode": "sorl",
+  "device": "cuda",
+  "push_to_hub": true,
+  "no_wandb": false,
+  "n_params": 162491102,
+  "run_name": "add_sub_sorl_v1_abs2_K1_10K",
+  "git_commit": "f835493c19eb98267697007042c9d440cad2afbb",
+  "timestamp": "2026-04-16T04:13:31.816099+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": "0zygf9qm",
+  "wandb_url": "https://wandb.ai/nlp_and_interpretability/sorl-arithmetic/runs/0zygf9qm",
+  "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": "5e90800683ec",
+  "final_accuracy": 0.9919230769230769,
+  "sft_accuracy": 0.9461538461538461,
+  "eval_method": "ArithmeticEvaluator"
+}