The previous step in research/lowcomm-training left two open points: the best inner learning rate for short sync intervals sat at the top of the grid, and outer momentum was switched off. Milestone 2f closes both on the same 14k-parameter character model (four workers, 640 steps each, five report seeds).
Adding an inner lr of 0.1 changed nothing: it loses at every sync interval H (and diverges for synchronous AdamW), so the earlier picks of 0.03 and 0.01 are real optima.
With the inner lr tuned for each setting, held-out loss at H = 16 / 64 / 160:
| outer step | H=16 | H=64 | H=160 |
|---|---|---|---|
| plain averaging | 2.183 | 2.210 | 2.218 |
| lr 1, momentum 0.5 | 2.191 | 2.200 | 2.269 |
| lr 0.7, momentum 0.9 (DiLoCo paper) | 2.218 | 2.257 | 2.486 |
Synchronous AdamW is 2.159. Momentum never closes the large-H gap; at H = 160 the paper's setting is 0.27 nats worse than plain averaging, with a seed sd of 0.07. The one small gain (H = 64, momentum 0.5) is about one standard deviation and is not claimed.
Limits: a model this small and under-trained says nothing about scale, only three outer settings were tried, and two seeds did the tuning. The log entry (hypotheses written before the run) and the results file are in the programme directory.