| { | |
| "dataset": { | |
| "cache_shards": 16, | |
| "data_dir": "data/flow_matching_10k", | |
| "include_hard_negatives": true, | |
| "max_shards": null, | |
| "max_source_weight": 5.0, | |
| "min_source_weight": 0.0, | |
| "normalize_trace": true, | |
| "seed": 42, | |
| "source_mode": "all", | |
| "split": "train", | |
| "split_file": "splits.json", | |
| "trace_transform": "sqrt", | |
| "use_source_weights": true, | |
| "validation_fraction": 0.1 | |
| }, | |
| "loss": { | |
| "path_noise_std": 0.0, | |
| "time_eps": 0.0001 | |
| }, | |
| "model": { | |
| "condition_dim": 256, | |
| "condition_on_source": false, | |
| "dropout": 0.0, | |
| "field_length": 256, | |
| "hidden_channels": 128, | |
| "residual_blocks": 8, | |
| "time_embedding_dim": 128, | |
| "trace_base_channels": 24, | |
| "trace_layers": 4 | |
| }, | |
| "training": { | |
| "amp": true, | |
| "batch_size": 64, | |
| "checkpoint_every": 1, | |
| "compile_model": false, | |
| "device": "cuda", | |
| "epochs": 50, | |
| "grad_clip_norm": 1.0, | |
| "learning_rate": 0.0003, | |
| "log_every": 25, | |
| "max_train_batches": null, | |
| "max_val_batches": null, | |
| "num_workers": 4, | |
| "out_dir": "runs/copra-flow", | |
| "seed": 42, | |
| "validate_every": 1, | |
| "weight_decay": 0.0001 | |
| } | |
| } | |