Add new #1 top-performing smooth-GmP model

Fine-tuned with GmP on COCO-SPRIGHT-40k as before, BUT with new a custom loss function and label smoothing: This CLIP has >91% accuracy on ImageNet/ObjectNet!

ViT-L-14-BEST-smooth-GmP-ft-pickle-OpenAI.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b2135fa3c6d8f9a64e3114b673e12ca38d50c77da0204ffeb34a9aeb29a7791e
+size 932307905

ViT-L-14-BEST-smooth-GmP-ft-state_dict.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:61661af02a59b5d5a175a94cba447ec5cc2318532ef202343a3be063b20981a9
+size 932240142

ViT-L-14-BEST-smooth-GmP-ft.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9fb57a8045c00fd501d9195515e94c0ae339d0fb9f46dc76cb20e685886494d9
+size 932103308