Merge pull request #85 from vitalops/batch_prefix_fix

Batch prefix fix

Author: fadil4u

datatune/core/filter.py

@@ -63,8 +63,9 @@ def llm_batch_inference(
     prompt = f"FILTERING CRITERIA:{os.linesep}{prompt}{os.linesep}{os.linesep}"
     suffix = (
         f"{os.linesep}{os.linesep}"
-        f"DECISION:Your response MUST be a valid Python dictionary in the format: {{key1: value1, key2: value2, ...}} with added key called '__filter__' with value either True to KEEP the record or False to REMOVE it."
+        f"DECISION:Your response MUST be the entire input record as  Python dictionary in the format: index=<row_index>|{{key1: value1, key2: value2, ...}}<endofrow> with added key called '__filter__' with value either True to KEEP the record or False to REMOVE it."
         f"No explanations or additional text."
+        f"ALWAYS STICK TO THE FORMAT index=<row_index>|{{key1: value1, key2: value2, ...}}<endofrow> with added key called '__filter__' with value either True to KEEP the record or False to REMOVE it."
     )
     df[llm_output_column] = llm(df[serialized_input_column], prefix, prompt, suffix, optimized=True)
     return df

datatune/core/map.py

@@ -70,6 +70,7 @@ def llm_batch_inference(
         f"""Your response MUST be the entire input record as a valid Python dictionary in the format
          'index=<row_index>|{{key1: value1, key2: value2, ...}}'  with added keys of expected new fields if any.
         ALWAYS START YOUR RESPONSE WITH 'index=<row_index>|' WHERE <row_index> IS THE INDEX OF THE ROW.
+        ALWAYS END YOUR RESPONSE WITH '<endofrow>'.
         """
     )
 

datatune/llm/llm.py

@@ -74,22 +74,16 @@ def _create_batched_prompts(
 
         """
         batch_prefix = (
-            "You will be given multiple data rows to process. Each request will:\n"
-            "- have the format 'index=<row_index>|{<row_data>}' where <row_index> is the zero-based index of the row in the original input list.\n"
-            "- End with '<endofrow>'\n\n"
-            "You MUST respond to each row in order. Each answer:\n"
-            "MUST BE OF THE FORMAT 'index=<row_index>|{response}<endofrow>' where <row_index> is the zero-based index of the row in the original input list.\n" \
-            "{response} must be enclosed in curly braces and strings should be enclosed in quotes.\n"
-            "- End with '<endofrow>'\n" \
-            "Always begin your response with 'index=<row_index>|' to indicate which row you are responding to without exception.\n"
-            "- Do NOT skip or omit any rows\n"
-            "Your entire response MUST include one answer per row. Respond strictly in the format described WITHOUT ANY OTHER TEXT, EXPLANATIONS OR BACKSTICKS\n" \
-            "IF DATA ROWS ARE NOT GIVEN IN DICTIONARY FORMAT RETURN THE ANSWER ONLY STARTING WITH index=<row_index>|"
-            "ALL RESPONSES MUST START WITH 'index='\n"
-            "ALL RESPONSES MUST END WITH '<endofrow>'\n"
-            f"Instructions:\n{batch_prefix or ''}"
+        "You will be given requests in the format 'index=<index>|{prompt}'. Each request ends with '<endofrow>'.\n"
+        "Respond to each prompt in order.\n"
+        "Each answer must be formatted exactly as 'index=<index>|{answer}<endofrow>'.\n"
+        "{answer} MUST BE ENCLOSED IN CURLY BRACES with strings in quotes.\n"
+        "DO NOT skip or omit any rows. DO NOT add explanations, backticks, or extra text.\n"
+        "Always start with 'index=<index>|' and end with '<endofrow>'.\n"
+        f"Instructions:\n{batch_prefix or ''}"
         )
 
+
         max_tokens = self.get_max_tokens()
         model_name = self._base_model_name
 
@@ -99,7 +93,7 @@ def _create_batched_prompts(
         nrows_per_api_call = []
         count = 0
         message = lambda x: [
-            {"role": "user", "content": f"{batch_prefix or ''}{x}{batch_suffix or ''}"}
+            {"role": "user", "content": f"{batch_prefix}{x}{batch_suffix or ''}"}
         ]
         prefix_suffix_tokens = token_counter(model_name, messages=message(""))
         total_ntokens = prefix_suffix_tokens