Add vote-level ARC evaluation artifacts

.gitattributes

@@ -58,3 +58,8 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 # Video files - compressed
 *.mp4 filter=lfs diff=lfs merge=lfs -text
 *.webm filter=lfs diff=lfs merge=lfs -text
+artifacts/arc_public_eval_aug_no_ttt_50m_3b_vllm_20260808/merged/predictions.jsonl filter=lfs diff=lfs merge=lfs -text
+artifacts/arc_public_eval_ttt_50m_3b_aggressive_hvote_20260808/merged/predictions.jsonl filter=lfs diff=lfs merge=lfs -text
+artifacts/arc_public_eval_ttt_50m_3b_replica_seed58_full400_20260809/predictions.jsonl filter=lfs diff=lfs merge=lfs -text
+artifacts/arc_public_eval_ttt_50m_3b_replica_seed59_full400_20260809/predictions.jsonl filter=lfs diff=lfs merge=lfs -text
+artifacts/arc_public_eval_ttt_7m_3b_aggressive_hvote_seed57_20260810/merged/predictions.jsonl filter=lfs diff=lfs merge=lfs -text

artifacts/arc_public_eval_aug_no_ttt_50m_3b_vllm_20260808/merged/predictions.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3420b6f8c359aebc5fa76319199f6882e2ad7810aaca7c86dbbfcb2f20a48330
+size 82079756

artifacts/arc_public_eval_aug_no_ttt_50m_3b_vllm_20260808/merged/summary.json

@@ -0,0 +1,97 @@
+{
+  "schema_version": 2,
+  "method": "frozen pretrained model with aggressive augmented inference",
+  "backend": "vllm",
+  "model_sha256": "5fd2d8c8f320be3debec07a1008d38e038e3c3401fdb292255dfc0b48bc77b43",
+  "evaluator_sha256": "0539fde0b88be845e3285204a449ea7600ced02ebdc39ea335b8a6fc916895a9",
+  "seed": 57,
+  "candidate_scoring": "original hierarchical transformed-vote top-2",
+  "accuracy_aggregation": "item-micro plus task-macro: mean query accuracy within task, then mean tasks",
+  "nested_candidate_budgets": [
+    14,
+    32,
+    64,
+    128
+  ],
+  "weights_updated": false,
+  "shard_count": 2,
+  "logical_candidate_predictions": 53632,
+  "unique_prompt_generations": 43053,
+  "valid_unique_generations": 43018,
+  "invalid_unique_generations": 35,
+  "aggregate_generation_seconds": 3358.691802352929,
+  "parallel_generation_wall_seconds": 1744.2013866659254,
+  "identity_prediction_matches_baseline": 408,
+  "identity_generated_ids_match_baseline": 408,
+  "identity_exact_labels_match_baseline": 419,
+  "identity_exact_gains_vs_baseline": 0,
+  "identity_exact_losses_vs_baseline": 0,
+  "identity_baseline_items": 419,
+  "task_count": 400,
+  "test_items": 419,
+  "identity_exact_items": 38,
+  "identity_item_accuracy": 0.09069212410501193,
+  "identity_task_macro_accuracy": 0.09,
+  "budgets": {
+    "14": {
+      "nominal_candidate_budget": 14,
+      "rank1_exact_items": 42,
+      "rank1_item_accuracy": 0.10023866348448687,
+      "rank1_task_macro_accuracy": 0.0975,
+      "top2_exact_items": 58,
+      "top2_item_accuracy": 0.13842482100238662,
+      "top2_task_macro_accuracy": 0.13375,
+      "top2_strict_tasks": 53,
+      "oracle_exact_items": 64,
+      "oracle_item_accuracy": 0.15274463007159905,
+      "oracle_task_macro_accuracy": 0.1475,
+      "mean_valid_candidates": 13.988066825775656,
+      "mean_distinct_predictions": 6.405727923627685
+    },
+    "32": {
+      "nominal_candidate_budget": 32,
+      "rank1_exact_items": 48,
+      "rank1_item_accuracy": 0.11455847255369929,
+      "rank1_task_macro_accuracy": 0.11125,
+      "top2_exact_items": 62,
+      "top2_item_accuracy": 0.14797136038186157,
+      "top2_task_macro_accuracy": 0.1425,
+      "top2_strict_tasks": 57,
+      "oracle_exact_items": 77,
+      "oracle_item_accuracy": 0.18377088305489261,
+      "oracle_task_macro_accuracy": 0.1775,
+      "mean_valid_candidates": 31.971360381861576,
+      "mean_distinct_predictions": 11.682577565632458
+    },
+    "64": {
+      "nominal_candidate_budget": 64,
+      "rank1_exact_items": 51,
+      "rank1_item_accuracy": 0.12171837708830549,
+      "rank1_task_macro_accuracy": 0.11875,
+      "top2_exact_items": 59,
+      "top2_item_accuracy": 0.14081145584725538,
+      "top2_task_macro_accuracy": 0.13625,
+      "top2_strict_tasks": 54,
+      "oracle_exact_items": 81,
+      "oracle_item_accuracy": 0.19331742243436753,
+      "oracle_task_macro_accuracy": 0.1875,
+      "mean_valid_candidates": 63.94749403341289,
+      "mean_distinct_predictions": 13.904534606205251
+    },
+    "128": {
+      "nominal_candidate_budget": 128,
+      "rank1_exact_items": 50,
+      "rank1_item_accuracy": 0.11933174224343675,
+      "rank1_task_macro_accuracy": 0.11625,
+      "top2_exact_items": 59,
+      "top2_item_accuracy": 0.14081145584725538,
+      "top2_task_macro_accuracy": 0.13625,
+      "top2_strict_tasks": 54,
+      "oracle_exact_items": 84,
+      "oracle_item_accuracy": 0.20047732696897375,
+      "oracle_task_macro_accuracy": 0.19375,
+      "mean_valid_candidates": 127.8926014319809,
+      "mean_distinct_predictions": 14.13365155131265
+    }
+  }
+}

artifacts/arc_public_eval_ttt_50m_3b_aggressive_hvote_20260808/merged/predictions.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:12c11abfc62b56512646d3880e0485c51dfa64e6d3ea7650121388562dd64701
+size 78507235

artifacts/arc_public_eval_ttt_50m_3b_aggressive_hvote_20260808/merged/summary.json

@@ -0,0 +1,127 @@
+{
+  "schema_version": 2,
+  "evaluator_sha256": "d04ad3340f02cb73afc759114154f18d8a20f02fbf4bb2c8cf5951c2dc986c24",
+  "method": "task-specific full-model TTT with aggressive augmented inference",
+  "training_reference": "Akyurek et al., arXiv:2411.07279v2",
+  "candidate_scoring": "original hierarchical transformed-vote top-2",
+  "accuracy_aggregation": "item-micro plus TRM-comparable task-macro: mean query accuracy within task, then mean tasks",
+  "model_path": "/home/nathanbreslow/arc/spark-gpt/checkpoints/arc_rearc_50m_3kpertask_seed0_20260807/hf_step0030516",
+  "model_sha256": "5fd2d8c8f320be3debec07a1008d38e038e3c3401fdb292255dfc0b48bc77b43",
+  "data_dir": "/home/nathanbreslow/arc/data/arc-agi-evaluation",
+  "task_count": 400,
+  "shard_index": null,
+  "shard_count": 2,
+  "seed": 57,
+  "max_examples": 250,
+  "max_seq_length": 16384,
+  "epochs": 2,
+  "batch_size": 2,
+  "optimizer": "MuonAdamWHybrid",
+  "muon_lr": 0.004,
+  "adamw_lr": 0.0005,
+  "muon_momentum": 0.95,
+  "weight_decay": 0.01,
+  "warmup_steps": 10,
+  "lr_decay_factor": 0.1,
+  "task_initialization": "exact restore of immutable pretrained checkpoint snapshot",
+  "model_compute_dtype": "bfloat16",
+  "loss": "output grid tokens after first demonstration output",
+  "inference_spatial_transformations": [
+    "identity",
+    "rotate90",
+    "rotate180",
+    "rotate270",
+    "flip_ud",
+    "flip_lr",
+    "transpose",
+    "anti_transpose"
+  ],
+  "inference_color_variants": 4,
+  "inference_demo_order_slots": 4,
+  "nested_candidate_budgets": [
+    14,
+    32,
+    64,
+    128
+  ],
+  "max_new_tokens": 1024,
+  "generation_batch_size": 16,
+  "deduplicate_identical_prompts_after_legacy_control": true,
+  "greedy": true,
+  "completed": true,
+  "trainable_parameters": 50372096,
+  "test_items": 419,
+  "unique_logical_prompts": 43053,
+  "actual_candidate_generations": 43572,
+  "logical_candidate_predictions": 53632,
+  "baseline_exact_on_scored_items": 35,
+  "baseline_item_accuracy_on_scored_items": 0.08353221957040573,
+  "budgets": {
+    "14": {
+      "nominal_candidate_budget": 14,
+      "rank1_exact_items": 155,
+      "rank1_item_accuracy": 0.36992840095465396,
+      "rank1_task_macro_accuracy": 0.3625,
+      "top2_exact_items": 187,
+      "top2_item_accuracy": 0.44630071599045346,
+      "top2_task_macro_accuracy": 0.43875,
+      "top2_strict_tasks": 175,
+      "oracle_exact_items": 221,
+      "oracle_item_accuracy": 0.5274463007159904,
+      "oracle_task_macro_accuracy": 0.52,
+      "mean_valid_candidates": 13.976133651551313,
+      "mean_distinct_predictions": 6.241050119331742
+    },
+    "32": {
+      "nominal_candidate_budget": 32,
+      "rank1_exact_items": 165,
+      "rank1_item_accuracy": 0.3937947494033413,
+      "rank1_task_macro_accuracy": 0.39,
+      "top2_exact_items": 193,
+      "top2_item_accuracy": 0.4606205250596659,
+      "top2_task_macro_accuracy": 0.45375,
+      "top2_strict_tasks": 180,
+      "oracle_exact_items": 240,
+      "oracle_item_accuracy": 0.5727923627684964,
+      "oracle_task_macro_accuracy": 0.5625,
+      "mean_valid_candidates": 31.96420047732697,
+      "mean_distinct_predictions": 11.536992840095465
+    },
+    "64": {
+      "nominal_candidate_budget": 64,
+      "rank1_exact_items": 165,
+      "rank1_item_accuracy": 0.3937947494033413,
+      "rank1_task_macro_accuracy": 0.39125,
+      "top2_exact_items": 199,
+      "top2_item_accuracy": 0.47494033412887826,
+      "top2_task_macro_accuracy": 0.47,
+      "top2_strict_tasks": 186,
+      "oracle_exact_items": 253,
+      "oracle_item_accuracy": 0.60381861575179,
+      "oracle_task_macro_accuracy": 0.595,
+      "mean_valid_candidates": 63.94510739856802,
+      "mean_distinct_predictions": 13.902147971360382
+    },
+    "128": {
+      "nominal_candidate_budget": 128,
+      "rank1_exact_items": 164,
+      "rank1_item_accuracy": 0.3914081145584726,
+      "rank1_task_macro_accuracy": 0.38875,
+      "top2_exact_items": 200,
+      "top2_item_accuracy": 0.477326968973747,
+      "top2_task_macro_accuracy": 0.4725,
+      "top2_strict_tasks": 187,
+      "oracle_exact_items": 259,
+      "oracle_item_accuracy": 0.6181384248210023,
+      "oracle_task_macro_accuracy": 0.6075,
+      "mean_valid_candidates": 127.89737470167064,
+      "mean_distinct_predictions": 14.262529832935561
+    }
+  },
+  "training_seconds": 12503.557789784398,
+  "generation_seconds": 15557.925329780803,
+  "shard_dirs": [
+    "/home/nathanbreslow/arc/evaluations/ttt_50m_3b_aggressive_hvote_shard0of2_20260808",
+    "/home/nathanbreslow/arc/evaluations/ttt_50m_3b_aggressive_hvote_shard1of2_20260808_merge_snapshot"
+  ]
+}

artifacts/arc_public_eval_ttt_50m_3b_replica_seed58_full400_20260809/predictions.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3e059b42f9edafacacd193171d557afd5d9624a0cd14424bda8d7cb03806e0ca
+size 78776070

artifacts/arc_public_eval_ttt_50m_3b_replica_seed58_full400_20260809/summary.json

@@ -0,0 +1,126 @@
+{
+  "schema_version": 2,
+  "evaluator_sha256": "af21b09d0b2d66fa0f5203498ec59fc7cd4ea346ff66f0805a48b1f2571992a3",
+  "method": "task-specific full-model TTT with aggressive augmented inference",
+  "training_reference": "Akyurek et al., arXiv:2411.07279v2",
+  "candidate_scoring": "original hierarchical transformed-vote top-2",
+  "accuracy_aggregation": "item-micro plus TRM-comparable task-macro: mean query accuracy within task, then mean tasks",
+  "model_path": "/home/nathanbreslow/arc/spark-gpt/checkpoints/arc_rearc_50m_3kpertask_seed0_20260807/hf_step0030516",
+  "model_sha256": "5fd2d8c8f320be3debec07a1008d38e038e3c3401fdb292255dfc0b48bc77b43",
+  "data_dir": "/home/nathanbreslow/arc/data/arc-agi-evaluation",
+  "task_count": 400,
+  "shard_index": 0,
+  "shard_count": 1,
+  "seed": 58,
+  "max_examples": 250,
+  "ttt_sampler": "legacy",
+  "diverse_sampler_sha256": null,
+  "diverse_max_attempts": 50000,
+  "max_seq_length": 16384,
+  "epochs": 2,
+  "batch_size": 2,
+  "optimizer": "MuonAdamWHybrid",
+  "muon_lr": 0.004,
+  "adamw_lr": 0.0005,
+  "muon_momentum": 0.95,
+  "weight_decay": 0.01,
+  "warmup_steps": 10,
+  "lr_decay_factor": 0.1,
+  "task_initialization": "exact restore of immutable pretrained checkpoint snapshot",
+  "model_compute_dtype": "bfloat16",
+  "loss": "output grid tokens after first demonstration output",
+  "inference_spatial_transformations": [
+    "identity",
+    "rotate90",
+    "rotate180",
+    "rotate270",
+    "flip_ud",
+    "flip_lr",
+    "transpose",
+    "anti_transpose"
+  ],
+  "inference_color_variants": 4,
+  "inference_demo_order_slots": 4,
+  "nested_candidate_budgets": [
+    14,
+    32,
+    64,
+    128
+  ],
+  "max_new_tokens": 1024,
+  "generation_batch_size": 16,
+  "deduplicate_identical_prompts_after_legacy_control": true,
+  "greedy": true,
+  "completed": true,
+  "trainable_parameters": 50372096,
+  "test_items": 419,
+  "unique_logical_prompts": 42910,
+  "actual_candidate_generations": 43388,
+  "logical_candidate_predictions": 53632,
+  "baseline_exact_on_scored_items": 0,
+  "baseline_item_accuracy_on_scored_items": null,
+  "budgets": {
+    "14": {
+      "nominal_candidate_budget": 14,
+      "rank1_exact_items": 155,
+      "rank1_item_accuracy": 0.36992840095465396,
+      "rank1_task_macro_accuracy": 0.3675,
+      "top2_exact_items": 180,
+      "top2_item_accuracy": 0.4295942720763723,
+      "top2_task_macro_accuracy": 0.42,
+      "top2_strict_tasks": 166,
+      "oracle_exact_items": 219,
+      "oracle_item_accuracy": 0.522673031026253,
+      "oracle_task_macro_accuracy": 0.51125,
+      "mean_valid_candidates": 13.99761336515513,
+      "mean_distinct_predictions": 6.31980906921241
+    },
+    "32": {
+      "nominal_candidate_budget": 32,
+      "rank1_exact_items": 159,
+      "rank1_item_accuracy": 0.3794749403341289,
+      "rank1_task_macro_accuracy": 0.37875,
+      "top2_exact_items": 192,
+      "top2_item_accuracy": 0.45823389021479716,
+      "top2_task_macro_accuracy": 0.45125,
+      "top2_strict_tasks": 179,
+      "oracle_exact_items": 232,
+      "oracle_item_accuracy": 0.5536992840095465,
+      "oracle_task_macro_accuracy": 0.5425,
+      "mean_valid_candidates": 31.98329355608592,
+      "mean_distinct_predictions": 11.799522673031026
+    },
+    "64": {
+      "nominal_candidate_budget": 64,
+      "rank1_exact_items": 176,
+      "rank1_item_accuracy": 0.4200477326968974,
+      "rank1_task_macro_accuracy": 0.4175,
+      "top2_exact_items": 203,
+      "top2_item_accuracy": 0.48448687350835323,
+      "top2_task_macro_accuracy": 0.47875,
+      "top2_strict_tasks": 190,
+      "oracle_exact_items": 244,
+      "oracle_item_accuracy": 0.5823389021479713,
+      "oracle_task_macro_accuracy": 0.57125,
+      "mean_valid_candidates": 63.957040572792366,
+      "mean_distinct_predictions": 13.968973747016706
+    },
+    "128": {
+      "nominal_candidate_budget": 128,
+      "rank1_exact_items": 180,
+      "rank1_item_accuracy": 0.4295942720763723,
+      "rank1_task_macro_accuracy": 0.4275,
+      "top2_exact_items": 204,
+      "top2_item_accuracy": 0.48687350835322196,
+      "top2_task_macro_accuracy": 0.4825,
+      "top2_strict_tasks": 191,
+      "oracle_exact_items": 248,
+      "oracle_item_accuracy": 0.5918854415274463,
+      "oracle_task_macro_accuracy": 0.58,
+      "mean_valid_candidates": 127.90930787589498,
+      "mean_distinct_predictions": 14.276849642004773
+    }
+  },
+  "training_seconds": 12826.654286775272,
+  "generation_seconds": 16064.380759360967
+}

artifacts/arc_public_eval_ttt_50m_3b_replica_seed59_full400_20260809/predictions.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a401db624ed1e73e95132550dd7691a0e85c52f683ee4d32471d8a553689eda6
+size 78450107

artifacts/arc_public_eval_ttt_50m_3b_replica_seed59_full400_20260809/summary.json

@@ -0,0 +1,126 @@
+{
+  "schema_version": 2,
+  "evaluator_sha256": "af21b09d0b2d66fa0f5203498ec59fc7cd4ea346ff66f0805a48b1f2571992a3",
+  "method": "task-specific full-model TTT with aggressive augmented inference",
+  "training_reference": "Akyurek et al., arXiv:2411.07279v2",
+  "candidate_scoring": "original hierarchical transformed-vote top-2",
+  "accuracy_aggregation": "item-micro plus TRM-comparable task-macro: mean query accuracy within task, then mean tasks",
+  "model_path": "/home/nathanbreslow/arc/spark-gpt/checkpoints/arc_rearc_50m_3kpertask_seed0_20260807/hf_step0030516",
+  "model_sha256": "5fd2d8c8f320be3debec07a1008d38e038e3c3401fdb292255dfc0b48bc77b43",
+  "data_dir": "/home/nathanbreslow/arc/data/arc-agi-evaluation",
+  "task_count": 400,
+  "shard_index": 0,
+  "shard_count": 1,
+  "seed": 59,
+  "max_examples": 250,
+  "ttt_sampler": "legacy",
+  "diverse_sampler_sha256": null,
+  "diverse_max_attempts": 50000,
+  "max_seq_length": 16384,
+  "epochs": 2,
+  "batch_size": 2,
+  "optimizer": "MuonAdamWHybrid",
+  "muon_lr": 0.004,
+  "adamw_lr": 0.0005,
+  "muon_momentum": 0.95,
+  "weight_decay": 0.01,
+  "warmup_steps": 10,
+  "lr_decay_factor": 0.1,
+  "task_initialization": "exact restore of immutable pretrained checkpoint snapshot",
+  "model_compute_dtype": "bfloat16",
+  "loss": "output grid tokens after first demonstration output",
+  "inference_spatial_transformations": [
+    "identity",
+    "rotate90",
+    "rotate180",
+    "rotate270",
+    "flip_ud",
+    "flip_lr",
+    "transpose",
+    "anti_transpose"
+  ],
+  "inference_color_variants": 4,
+  "inference_demo_order_slots": 4,
+  "nested_candidate_budgets": [
+    14,
+    32,
+    64,
+    128
+  ],
+  "max_new_tokens": 1024,
+  "generation_batch_size": 16,
+  "deduplicate_identical_prompts_after_legacy_control": true,
+  "greedy": true,
+  "completed": true,
+  "trainable_parameters": 50372096,
+  "test_items": 419,
+  "unique_logical_prompts": 42973,
+  "actual_candidate_generations": 43452,
+  "logical_candidate_predictions": 53632,
+  "baseline_exact_on_scored_items": 0,
+  "baseline_item_accuracy_on_scored_items": null,
+  "budgets": {
+    "14": {
+      "nominal_candidate_budget": 14,
+      "rank1_exact_items": 147,
+      "rank1_item_accuracy": 0.35083532219570407,
+      "rank1_task_macro_accuracy": 0.34375,
+      "top2_exact_items": 174,
+      "top2_item_accuracy": 0.4152744630071599,
+      "top2_task_macro_accuracy": 0.4025,
+      "top2_strict_tasks": 159,
+      "oracle_exact_items": 218,
+      "oracle_item_accuracy": 0.5202863961813843,
+      "oracle_task_macro_accuracy": 0.5075,
+      "mean_valid_candidates": 13.985680190930788,
+      "mean_distinct_predictions": 6.317422434367542
+    },
+    "32": {
+      "nominal_candidate_budget": 32,
+      "rank1_exact_items": 158,
+      "rank1_item_accuracy": 0.37708830548926014,
+      "rank1_task_macro_accuracy": 0.3725,
+      "top2_exact_items": 194,
+      "top2_item_accuracy": 0.4630071599045346,
+      "top2_task_macro_accuracy": 0.4525,
+      "top2_strict_tasks": 179,
+      "oracle_exact_items": 239,
+      "oracle_item_accuracy": 0.5704057279236276,
+      "oracle_task_macro_accuracy": 0.55875,
+      "mean_valid_candidates": 31.94033412887828,
+      "mean_distinct_predictions": 11.599045346062052
+    },
+    "64": {
+      "nominal_candidate_budget": 64,
+      "rank1_exact_items": 164,
+      "rank1_item_accuracy": 0.3914081145584726,
+      "rank1_task_macro_accuracy": 0.38875,
+      "top2_exact_items": 205,
+      "top2_item_accuracy": 0.4892601431980907,
+      "top2_task_macro_accuracy": 0.48,
+      "top2_strict_tasks": 189,
+      "oracle_exact_items": 250,
+      "oracle_item_accuracy": 0.5966587112171837,
+      "oracle_task_macro_accuracy": 0.585,
+      "mean_valid_candidates": 63.89021479713604,
+      "mean_distinct_predictions": 13.928400954653938
+    },
+    "128": {
+      "nominal_candidate_budget": 128,
+      "rank1_exact_items": 159,
+      "rank1_item_accuracy": 0.3794749403341289,
+      "rank1_task_macro_accuracy": 0.37625,
+      "top2_exact_items": 201,
+      "top2_item_accuracy": 0.4797136038186158,
+      "top2_task_macro_accuracy": 0.4725,
+      "top2_strict_tasks": 186,
+      "oracle_exact_items": 255,
+      "oracle_item_accuracy": 0.6085918854415274,
+      "oracle_task_macro_accuracy": 0.5975,
+      "mean_valid_candidates": 127.78520286396181,
+      "mean_distinct_predictions": 14.16706443914081
+    }
+  },
+  "training_seconds": 12526.061385635985,
+  "generation_seconds": 15556.743899527792
+}

artifacts/arc_public_eval_ttt_7m_3b_aggressive_hvote_seed57_20260810/merged/predictions.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d3e787ceba9a5c9a77d4461d04dc5850ba01678e08f632e9aec736cf1a70e74c
+size 77668253

artifacts/arc_public_eval_ttt_7m_3b_aggressive_hvote_seed57_20260810/merged/summary.json

@@ -0,0 +1,130 @@
+{
+  "schema_version": 2,
+  "evaluator_sha256": "af21b09d0b2d66fa0f5203498ec59fc7cd4ea346ff66f0805a48b1f2571992a3",
+  "method": "task-specific full-model TTT with aggressive augmented inference",
+  "training_reference": "Akyurek et al., arXiv:2411.07279v2",
+  "candidate_scoring": "original hierarchical transformed-vote top-2",
+  "accuracy_aggregation": "item-micro plus TRM-comparable task-macro: mean query accuracy within task, then mean tasks",
+  "model_path": "/home/nathanbreslow/arc/spark-gpt/checkpoints/arc_rearc_7m_3kpertask_seed0_20260810/hf",
+  "model_sha256": "b5dd708baa7785727a9cb255a677c99c120809eb7bb2017e85181285accad524",
+  "data_dir": "/home/nathanbreslow/arc/data/arc-agi-evaluation",
+  "task_count": 400,
+  "shard_index": null,
+  "shard_count": 2,
+  "seed": 57,
+  "max_examples": 250,
+  "ttt_sampler": "legacy",
+  "diverse_sampler_sha256": null,
+  "diverse_max_attempts": 50000,
+  "max_seq_length": 16384,
+  "epochs": 2,
+  "batch_size": 2,
+  "optimizer": "MuonAdamWHybrid",
+  "muon_lr": 0.004,
+  "adamw_lr": 0.0005,
+  "muon_momentum": 0.95,
+  "weight_decay": 0.01,
+  "warmup_steps": 10,
+  "lr_decay_factor": 0.1,
+  "task_initialization": "exact restore of immutable pretrained checkpoint snapshot",
+  "model_compute_dtype": "bfloat16",
+  "loss": "output grid tokens after first demonstration output",
+  "inference_spatial_transformations": [
+    "identity",
+    "rotate90",
+    "rotate180",
+    "rotate270",
+    "flip_ud",
+    "flip_lr",
+    "transpose",
+    "anti_transpose"
+  ],
+  "inference_color_variants": 4,
+  "inference_demo_order_slots": 4,
+  "nested_candidate_budgets": [
+    14,
+    32,
+    64,
+    128
+  ],
+  "max_new_tokens": 1024,
+  "generation_batch_size": 16,
+  "deduplicate_identical_prompts_after_legacy_control": true,
+  "greedy": true,
+  "completed": true,
+  "trainable_parameters": 7094784,
+  "test_items": 419,
+  "unique_logical_prompts": 43053,
+  "actual_candidate_generations": 43572,
+  "logical_candidate_predictions": 53632,
+  "baseline_exact_on_scored_items": 0,
+  "baseline_item_accuracy_on_scored_items": null,
+  "budgets": {
+    "14": {
+      "nominal_candidate_budget": 14,
+      "rank1_exact_items": 104,
+      "rank1_item_accuracy": 0.24821002386634844,
+      "rank1_task_macro_accuracy": 0.24,
+      "top2_exact_items": 119,
+      "top2_item_accuracy": 0.2840095465393795,
+      "top2_task_macro_accuracy": 0.2725,
+      "top2_strict_tasks": 108,
+      "oracle_exact_items": 157,
+      "oracle_item_accuracy": 0.3747016706443914,
+      "oracle_task_macro_accuracy": 0.36,
+      "mean_valid_candidates": 13.79236276849642,
+      "mean_distinct_predictions": 7.22673031026253
+    },
+    "32": {
+      "nominal_candidate_budget": 32,
+      "rank1_exact_items": 120,
+      "rank1_item_accuracy": 0.2863961813842482,
+      "rank1_task_macro_accuracy": 0.28,
+      "top2_exact_items": 142,
+      "top2_item_accuracy": 0.33890214797136037,
+      "top2_task_macro_accuracy": 0.32875,
+      "top2_strict_tasks": 130,
+      "oracle_exact_items": 194,
+      "oracle_item_accuracy": 0.4630071599045346,
+      "oracle_task_macro_accuracy": 0.45,
+      "mean_valid_candidates": 31.54653937947494,
+      "mean_distinct_predictions": 13.976133651551313
+    },
+    "64": {
+      "nominal_candidate_budget": 64,
+      "rank1_exact_items": 117,
+      "rank1_item_accuracy": 0.27923627684964203,
+      "rank1_task_macro_accuracy": 0.275,
+      "top2_exact_items": 145,
+      "top2_item_accuracy": 0.3460620525059666,
+      "top2_task_macro_accuracy": 0.3375,
+      "top2_strict_tasks": 134,
+      "oracle_exact_items": 211,
+      "oracle_item_accuracy": 0.5035799522673031,
+      "oracle_task_macro_accuracy": 0.49,
+      "mean_valid_candidates": 63.11217183770883,
+      "mean_distinct_predictions": 16.367541766109785
+    },
+    "128": {
+      "nominal_candidate_budget": 128,
+      "rank1_exact_items": 119,
+      "rank1_item_accuracy": 0.2840095465393795,
+      "rank1_task_macro_accuracy": 0.2775,
+      "top2_exact_items": 145,
+      "top2_item_accuracy": 0.3460620525059666,
+      "top2_task_macro_accuracy": 0.3375,
+      "top2_strict_tasks": 134,
+      "oracle_exact_items": 217,
+      "oracle_item_accuracy": 0.5178997613365155,
+      "oracle_task_macro_accuracy": 0.505,
+      "mean_valid_candidates": 126.14319809069212,
+      "mean_distinct_predictions": 16.608591885441527
+    }
+  },
+  "training_seconds": 3103.418425070413,
+  "generation_seconds": 4993.645875780901,
+  "shard_dirs": [
+    "/home/nathanbreslow/arc/evaluations/ttt_7m_3b_aggressive_hvote_seed57_shard0of2_20260810",
+    "/home/nathanbreslow/arc/evaluations/ttt_7m_3b_aggressive_hvote_seed57_shard1of2_20260810_from_bob"
+  ]
+}