Improve sweep defaults (#58)

Default to something that works with smaller GPUs. Use the L1 from SAE training guide.

Author: alan-cooney

sparse_autoencoder/train/sweep_config.py

@@ -75,9 +75,9 @@ class SweepParametersRuntime(dict[str, Any]):
 
     adam_weight_decay: float = 0.0
 
-    l1_coefficient: float = 0.01
+    l1_coefficient: float = 0.001
 
-    batch_size: int = 8192
+    batch_size: int = 512
 
     def to_dict(self) -> dict[str, Any]:
         """Return dict representation of this object."""