add --lr flag (wide-adapter LR rescale diagnostics at 12B)

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

scripts/train_merge_code.py

@@ -30,7 +30,6 @@ OUT = Path(os.environ.get("LOOP_OUT",
     Path(__file__).resolve().parent.parent / "results-loop"))
 STEPS = 600
 BATCH = 4
-LR = 1e-3
 WARMUP = 20
 MAX_TOK = 512
 K_BUCKETS = [(1, ("easy",)), (2, ("easy", "hard")), (4, ("hard",))]
@@ -51,13 +50,16 @@ ap.add_argument("--deepk", type=int, default=0,
                 help="scale curriculum depths by deepk/4 (e.g. 16 -> 2/8/16)")
 ap.add_argument("--bptt", type=int, default=0,
                 help="truncated BPTT: grads only through last N iterations")
+ap.add_argument("--lr", type=float, default=1e-3)
 ARGS = ap.parse_args()
 SEED = ARGS.seed
+LR = ARGS.lr
 SUFFIX = ((f"_s{SEED}" if SEED else "")
           + (f"_p{ARGS.pause}" if ARGS.pause else "")
           + (f"_a{ARGS.alpha}" if ARGS.alpha != 0.3 else "")
           + ("_ad" if ARGS.adaptive else "")
-          + (f"_dk{ARGS.deepk}" if ARGS.deepk else ""))
+          + (f"_dk{ARGS.deepk}" if ARGS.deepk else "")
+          + (f"_lr{ARGS.lr}" if ARGS.lr != 1e-3 else ""))
 PAUSE_ID = 6  # <unused0>