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 | 5,000 | checkpoint_step_00005000.pt |
| rawtoken_n2_k50_d50_seed0 | 5,000 | checkpoint_step_00005000.pt |
| multitoken_n2_k4_d100_seed0 | 10,000 | checkpoint_step_00010000.pt |
| rawtoken_n2_k50_d100_seed0 | 10,000 | checkpoint_step_00010000.pt |
| multitoken_n2_k4_d300_seed0 | 30,000 | checkpoint_step_00030000.pt |
| rawtoken_n2_k50_d300_seed0 | 30,000 | checkpoint_step_00030000.pt |
| multitoken_n2_k4_d1000_seed0 | 100,000 | checkpoint_step_00100000.pt |
| rawtoken_n2_k50_d1000_seed0 | 100,000 | checkpoint_step_00100000.pt |
| multitoken_n2_k4_d2900_seed0 | 280,000 | checkpoint_step_00280000.pt |
| rawtoken_n2_k50_d2900_seed0 | 280,000 | checkpoint_step_00280000.pt |
The K=1 controls reuse the N2 / seed-0 checkpoints in Section 4.