Added some split prompting started code, adamw8bit, replacements improving, learnable snr gos. A lot of good stuff.

extensions_built_in/dataset_tools/tools/caption.py

@@ -33,7 +33,13 @@ def clean_caption(cap, replacements=None):
     cap = " ".join(cap.split())
 
     for replacement in replacements:
-        cap = cap.replace(replacement[0], replacement[1])
+        if replacement[0].startswith('*'):
+            # we are removing all text if it starts with this and the rest matches
+            search_text = replacement[0][1:]
+            if cap.startswith(search_text):
+                cap = ""
+        else:
+            cap = cap.replace(replacement[0].lower(), replacement[1].lower())
 
     cap_list = cap.split(",")
     # trim whitespace

extensions_built_in/dataset_tools/tools/llava_utils.py

@@ -77,7 +77,7 @@ class LLaVAImageProcessor:
             output_ids = self.model.generate(
                 input_ids, images=image_tensor, do_sample=True, temperature=0.1,
                 max_new_tokens=max_new_tokens, use_cache=True, stopping_criteria=[stopping_criteria],
-                top_p=0.9
+                top_p=0.8
             )
         outputs = self.tokenizer.decode(output_ids[0, input_ids.shape[1]:]).strip()
         conv.messages[-1][-1] = outputs