Scope

Tiny docs-only change per #38 continuation agent's recommendation. No code change needed since default is already correct.

Change

In `configs/training/default.yaml`, add a comment near where `packing` could be added (or in a relevant note section):

```yaml

NOTE: packing defaults to False for the in-process LoRA path.

Pilot #38 (c1_evil_wrong_em, 2 seeds, 2 arms) showed packing=True on small

coupling datasets:

- slower (-10.5% tokens/sec)

- under-trains coupling signal (+28 pt alignment toward base, +0.27 train_loss)

- collapses optimizer steps 3.67× with same token budget

For distributed Tulu-scale SFT, packing=True is different — see configs/tulu/

configs. Revisit the LoRA default only for long-sequence realistic data.

```

Budget

5 min. No GPU.