Fix Auto Prompt pipeline in sandbox (#791)

* * involve answer of QAs for grader

* * update reference of flagged_words_filter

Author: HYLcool

configs/data_juicer_recipes/sandbox/auto_prompt_optimization/sandbox_auto_prompt_optimization.yaml

@@ -10,7 +10,7 @@ resume: true                                                              # allo
 # iteration related parameters
 max_iter_num: 5
 iter_targets:
-  - "grader_model_prompt_optimization.grades_evaluation.min_mse <= 0.26"
+  - "grader_model_prompt_optimization.grades_evaluation.min_mse <= 0.2"
 iter_updater:
   select_and_merge_data_pools.merge_single_prompt_data_pools.merged_top_prompt_dataset: grader_model_prompt_optimization.generate_new_prompts.dj_configs.dataset_path
 
@@ -53,11 +53,12 @@ pipelines:
             type: 'api'
             model: 'qwen2.5-32b-instruct'
             max_retry_num: 5
-            build_messages_func: 'build_messages'
+            build_messages_func: 'build_messages_for_math_qa'
             parse_output_func: 'parse_output'
             func_kwargs:
               system_key: "prompt"
               query_key: "question"
+              response_key: "answer"
             # data related
             dataset_path: '<replaced_by_the_input>'
             export_path: './outputs/auto-prompt-optimization-for-grader-model/infer_results/'

data_juicer/core/sandbox/helper_funcs.py

@@ -27,3 +27,23 @@ def parse_output(output: str, item: dict, **kwargs):
     A simple implementation.
     """
     return output
+
+
+# Math QA grader
+@ALL_FUNCS.register_module("build_messages_for_math_qa")
+def build_messages_for_math_qa(item: dict, **kwargs):
+    """
+    Build message for math QA grader.
+    """
+    system_key = kwargs.get("system_key", "system")
+    query_key = kwargs.get("query_key", "query")
+    response_key = kwargs.get("response_key", "response")
+
+    system_prompt = item.get(system_key, "")
+    question = item[query_key]
+    answer = item[response_key]
+    messages = []
+    if system_prompt:
+        messages.append({"role": "system", "content": system_prompt})
+    messages.append({"role": "user", "content": f"Question: {question}\nAnswer: {answer}"})
+    return messages

data_juicer/ops/filter/flagged_words_filter.py

@@ -1,5 +1,8 @@
 # Some code here has been modified from:
 # https://huggingface.co/spaces/huggingface/text-data-filtering
+#
+# The flagged words list comes from https://huggingface.co/spaces/huggingface/text-data-filtering
+# and https://github.com/LDNOOBW/List-of-Dirty-Naughty-Obscene-and-Otherwise-Bad-Words
 # --------------------------------------------------------
 
 from typing import List