| # Section 5 — Token interface | |
| Paper: **Do more tokens per timestep improve performance?** | |
| Ten final checkpoints: N2, K∈{4,50}, seed 0, C_train=10, across all five data settings. | |
| | Run | Training steps | Checkpoint file | | |
| |---|---:|---| | |
| | [multitoken_n2_k4_d50_seed0](multitoken_n2_k4_d50_seed0/) | 5,000 | `checkpoint_step_00005000.pt` | | |
| | [rawtoken_n2_k50_d50_seed0](rawtoken_n2_k50_d50_seed0/) | 5,000 | `checkpoint_step_00005000.pt` | | |
| | [multitoken_n2_k4_d100_seed0](multitoken_n2_k4_d100_seed0/) | 10,000 | `checkpoint_step_00010000.pt` | | |
| | [rawtoken_n2_k50_d100_seed0](rawtoken_n2_k50_d100_seed0/) | 10,000 | `checkpoint_step_00010000.pt` | | |
| | [multitoken_n2_k4_d300_seed0](multitoken_n2_k4_d300_seed0/) | 30,000 | `checkpoint_step_00030000.pt` | | |
| | [rawtoken_n2_k50_d300_seed0](rawtoken_n2_k50_d300_seed0/) | 30,000 | `checkpoint_step_00030000.pt` | | |
| | [multitoken_n2_k4_d1000_seed0](multitoken_n2_k4_d1000_seed0/) | 100,000 | `checkpoint_step_00100000.pt` | | |
| | [rawtoken_n2_k50_d1000_seed0](rawtoken_n2_k50_d1000_seed0/) | 100,000 | `checkpoint_step_00100000.pt` | | |
| | [multitoken_n2_k4_d2900_seed0](multitoken_n2_k4_d2900_seed0/) | 280,000 | `checkpoint_step_00280000.pt` | | |
| | [rawtoken_n2_k50_d2900_seed0](rawtoken_n2_k50_d2900_seed0/) | 280,000 | `checkpoint_step_00280000.pt` | | |
| The K=1 controls reuse the N2 / seed-0 checkpoints in [Section 4](../04_robocasa_scaling/README.md). | |