weird fix

.gitignore

@@ -2,3 +2,5 @@
 .Rhistory
 .RData
 .Ruserdata
+README.Rmd
+inst/doc

R/hollr.R

@@ -54,7 +54,7 @@ hollr <- function(id,
                   extract_json = TRUE) {
 
   # Determine if the model is OpenAI or local
-  is_openai_model <- grepl("gpt-3.5-turbo|gpt-4|gpt-4o|gpt-4o-mini", model, ignore.case = TRUE)
+  is_openai_model <- grepl("^gpt-", model, ignore.case = TRUE)
 
   if (!is_openai_model && batch_size > 1) {
     # Run the local model with batch processing
@@ -110,7 +110,7 @@ hollr <- function(id,
                                                force_json, 
                                                max_attempts)
 
-    cleaned_response <- gsub("^\\s*\\[\\s*\\{.*\\}\\s*\\]\\s*$", "", validation_result$response)
+    cleaned_response <- gsub("^```json\\s*|```$", "", validation_result$response)
 
     list(id = row$id,
          annotator_id = row$annotator_id,

inst/python/llm_functions.py

@@ -1,5 +1,3 @@
-
-
 def initialize_model(model_name):
     """
     Initialize a text generation model using the Hugging Face transformers library.
@@ -10,22 +8,18 @@ def initialize_model(model_name):
     Returns:
         pipeline: A text generation pipeline using the specified model and tokenizer.
     """
-    import os
-    import torch
     from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline
-
-    # Set the CUDA allocation configuration
-    os.environ['PYTORCH_CUDA_ALLOC_CONF'] = 'expandable_segments:True'
-
-    # Optionally, clear the CUDA cache before starting
-    torch.cuda.empty_cache()
+    import torch
 
     # Load the tokenizer for the specified model
     tokenizer = AutoTokenizer.from_pretrained(model_name)
+    print("Initial pad token:", tokenizer.pad_token)  # Debugging: Print initial pad token
 
     # Set pad token if it does not exist
     if tokenizer.pad_token is None:
+        print("No pad token found, adding '[PAD]' as pad_token")
         tokenizer.add_special_tokens({'pad_token': '[PAD]'})
+        print("New pad token set:", tokenizer.pad_token)
 
     # Load the model
     model = AutoModelForCausalLM.from_pretrained(model_name,
@@ -34,17 +28,18 @@ def initialize_model(model_name):
 
     # Update model embeddings
     model.resize_token_embeddings(len(tokenizer))
-
-    # Optionally, clear the CUDA cache after model initialization
-    torch.cuda.empty_cache()
+    print("Model embeddings updated to reflect new tokenizer size.")
 
     # Create text generation pipeline
     model_pipeline = pipeline("text-generation", model=model, tokenizer=tokenizer)
+    print("Pipeline created with pad token:", tokenizer.pad_token)  # Confirm pad token at pipeline creation
 
     return model_pipeline
 
 
 
+
+
 def generate_text_batch(model_pipeline, prompt, temperature, max_length, max_new_tokens=None):
     """
     Generate text in batch using the specified text generation pipeline, ensuring the input prompt is not included in the output.