Upload training_config.yml with huggingface_hub

training_config.yml

@@ -8,6 +8,7 @@ model:
   lora_rank: 8
   lora_alpha: 16
   perception_tokens: 2
+  use_clip: false
 tokenizer:
   _component_: models.a2a_tokenizer
   path: checkpoints/Meta-Llama-3-8B-Instruct/tokenizer.model
@@ -18,28 +19,21 @@ checkpointer:
   - consolidated.00.pth
   adapter_checkpoint: null
   recipe_checkpoint: null
-  output_dir: output
+  output_dir: output_checkpoints/experiment_1
   model_type: LLAMA3
 resume_from_checkpoint: false
+interim_checkpoint_steps: 5000
+interim_gen_steps: null
+max_new_tokens: 100
+temperature: 0.6
+top_k: 300
 dataset:
   _component_: ds.EvenBatcher
   dataset:
     _component_: ds.RoundRobinDataset
     datasets:
     - _component_: ds.LlavaInstructDataset
-      ib_embed_path: ds/coco_llava_instruct/*.ib_embed.pt
-      clip_embed_path: ds/coco_llava_instruct/*.clip_embed.pt
-      caption_path: ds/coco_llava_instruct/*.caption.pt
-      train_on_input: false
-    - _component_: ds.LlavaInstructDataset
-      ib_embed_path: ds/vision_flan/*.ib_embed.pt
-      clip_embed_path: ds/vision_flan/*.clip_embed.pt
-      caption_path: ds/vision_flan/*.caption.pt
-      train_on_input: false
-    - _component_: ds.CaptionInstructDataset
-      ib_embed_path: ds/sam_llava/00.ib_embed.pt
-      clip_embed_path: ds/sam_llava/00.clip_embed.pt
-      caption_path: ds/sam_llava/00.caption.pt
+      dataset_path: ds/coco_llava_instruct/output.parquet
       train_on_input: false
 seed: null
 shuffle: true