| { | |
| "model": "fusion-embedding-2-tactus", | |
| "version": "v0.1-preview", | |
| "ckpt_file": "tactus_head.pt", | |
| "grid": 32, | |
| "window_frames": 8, | |
| "depth": "resnet18", | |
| "temporal": "conv", | |
| "flow": false, | |
| "feat_dim": 512, | |
| "out_dim": 2048, | |
| "input_normalization": "stag_affine_500_650: clip((raw-500)/150, 0, 1); pass [0,1] floats or uint8 0-255 maps", | |
| "text_target": "canonical UnifiedEmbedder.embed_text (whitened, 2048-d) from EximiusLabs/fusion-embedding-2-2b-preview", | |
| "mae_init": "tactus_mae_resnet18_stagonly_equal_npx (144k STAG-family frames, mask 0.6, norm-pix targets)", | |
| "train_recipe": { | |
| "steps": 16000, | |
| "sampling": "cluster", | |
| "extra_data": [ | |
| "blindfolded", | |
| "weights" | |
| ], | |
| "augment": true, | |
| "schedule": "cosine", | |
| "temp": 0.07 | |
| } | |
| } |