Revert to 5a38762f (end of stage 1) — drop second stage 2 attempt

Second stage-2 attempt was launched with the milder aug recipe but still at the original peak LRs (lr=1e-3, decoder_lr=1e-4), which is too high for resumed-from-converged training. Lowering peak LRs (lr=3e-4, decoder_lr=3e-5) and restarting from the stage-1 baseline.

model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:606624b5ff277e957e379645f0fbe86631c5dca87e02da9c483da2d3fde5d869
+oid sha256:8e830a9e4362b32186ad529b33d2fb7c7c930ca182d471178f2d24bee05e3bb4
 size 2433494416

training_args.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:804d2e91db712a125d1f04c9d728582a8168f28597182b48feeddc16222814c1
+oid sha256:323f3fcb8fda7d037add78c657f2f95394b69825fe27e5f794b4b2607076e45a
 size 5329