benchmark: antipasto rotate_basis CLI + lr=5e-3 + ablation queue

justfile

@@ -88,6 +88,7 @@ bench-variant model variant steps="5000":
 		delora) lr=1e-3 ;;
 		ia3)    lr=5e-3; target='(k_proj|v_proj)$' ;;
 		ia3_ff) lr=5e-3; target='(down_proj)$' ;;
+		antipasto) lr=5e-3 ;;  # small params need higher lr
 	esac
 	exec uv run --extra benchmark python scripts/metamath_gsm8k_benchmark.py \
 		--model '{{model}}' \
@@ -95,7 +96,7 @@ bench-variant model variant steps="5000":
 		--steps {{steps}} \
 		--lr "$lr" \
 		--target-name "$target" \
-		--layers all --r 32 --alpha 64
+		--layers all --r 32 --alpha 64 "$@"
 
 metamath-queue-all model="Qwen/Qwen3-0.6B-Base" steps="5000" variants="lora pissa delora dora hra ia3 ia3_ff eva antipasto":
 	#!/usr/bin/env bash

scripts/metamath_gsm8k_benchmark.py

@@ -52,6 +52,7 @@ class BenchmarkConfig:
     alpha: float = 64.0
     delora_lambda0: float = 0.1
     road_group_size: int = 64
+    antipasto_rotate_basis: Literal["V", "U", "none"] = "V"
     target_name: list[str] = field(default_factory=lambda: list(DEFAULT_TARGETS))
     layers: str = "all"
     train_dataset: str = "meta-math/MetaMathQA"
@@ -123,6 +124,8 @@ def cfg_for_variant(args: BenchmarkConfig, dtype: torch.dtype) -> ll.AdapterConf
     extra = {"lambda0": args.delora_lambda0} if args.variant == "delora" else {}
     if args.variant == "road":
         extra = {"group_size": args.road_group_size}
+    if args.variant == "antipasto":
+        extra = {"rotate_basis": args.antipasto_rotate_basis}
     return CFG_BY_VARIANT[args.variant](
         r=args.r,
         alpha=args.r if args.variant == "pissa" else args.alpha,