Update merge.py

merge.py

@@ -17,8 +17,8 @@ model = PeftModel.from_pretrained(model, "/path/to/llama3-8b-adapter", device_ma
 model = model.merge_and_unload()
 
 tokenizer = AutoTokenizer.from_pretrained("/path/to/meta-llama3-8b", trust_remote_code=True)
-tokenizer.pad_token_id = 18610
-
+tokenizer.pad_token_id = tokenizer.eos_token_id + 1
+model.config.pad_token_id = tokenizer.pad_token_id
 pipe = pipeline(task="text-generation", model=model, tokenizer=tokenizer, max_length=4096, do_sample=False)
 print("Padding side:",tokenizer.padding_side)
 val_dataset = load_dataset("csv", data_files={'val':'/path/to/actseq-val-new.csv'})["val"]
@@ -28,7 +28,7 @@ test_dataset = load_dataset("csv", data_files={'test':'/path/to/actseq-test-new.
 def formatting_prompts_func(example):
      output_texts = []
      for i in range(len(example['dial_with_actions'])):
-         text = f"Predict the action sequence (AS) for the Minecraft excerpt:\n {example['dial_with_actions'][i]}\n ### AS:"
+         text = f"<|begin_of_text|>Predict the action sequence (AS) for the Minecraft excerpt:\n {example['dial_with_actions'][i]}\n ### AS:"
          output_texts.append(text)
      return output_texts