{ "data": { "arc_agi1_path": "References/ARC-AGI", "arc_agi2_path": "References/ARC-AGI-2", "re_arc_path": "References/RE-ARC", "canvas_size": 64, "num_colors": 12, "max_grid_size": 30, "max_demos": 5, "re_arc_samples_per_task": 50, "repeat_factor": 1, "augmentation": { "geometric": true, "color_permutation": true, "num_color_perms": 10, "keep_background": true, "resolution_scaling": true, "translation": true } }, "model": { "hidden_size": 384, "mlp_ratio": 2.5, "tokenizer": { "patch_size": 4 }, "rule_encoder": { "pair_layers": 2, "agg_layers": 2, "num_heads": 8, "num_rule_tokens": 64 }, "rule_applier": { "num_layers": 4, "num_heads": 8 }, "decoder": { "upsample_method": "transposed_conv", "hidden_channels": [ 192, 96 ] } }, "training": { "batch_size": 64, "num_workers": 8, "gradient_clip": 1.0, "stage1": { "name": "pretrain", "data_sources": [ "re_arc" ], "epochs": 50, "lr": 0.0003 }, "stage2": { "name": "finetune", "data_sources": [ "agi1", "agi2" ], "epochs": 30, "lr": 0.0001 }, "stage3": { "name": "hard_focus", "data_sources": [ "agi1", "agi2" ], "epochs": 10, "lr": 3e-05, "agi2_oversample": 2.0 }, "optimizer": { "name": "adamw", "weight_decay": 0.01, "betas": [ 0.9, 0.999 ] }, "scheduler": { "name": "cosine", "warmup_ratio": 0.1 }, "log_every_n_steps": 100, "save_every_n_epochs": 10, "checkpoint_dir": "checkpoints" }, "ttt": { "enabled": true, "steps": 100, "lr": 0.0001, "batch_size": 8, "num_candidates": 32 }, "evaluation": { "val_split_ratio": 0.1, "val_data_sources": [ "agi1", "agi2" ], "metrics": [ "pixel_accuracy", "grid_exact_match" ], "visualize_every_n_tasks": 50 } }