Merge pull request #163 from ls1intum/add-grading-critera-support-llm

Add grading criteria support for llm modules

Author: FelixTJDietrich

athena/athena/__init__.py

@@ -3,7 +3,7 @@
 from pathlib import Path
 
 from .app import app
-from .schemas import ExerciseType
+from .schemas import ExerciseType, GradingCriterion, StructuredGradingInstruction
 from .metadata import emit_meta, get_meta
 from .experiment import get_experiment_environment
 from .endpoints import submission_selector, submissions_consumer, feedback_consumer, feedback_provider, config_schema_provider  # type: ignore
@@ -32,5 +32,7 @@ def run_module():
     "get_meta",
     "get_experiment_environment",
     "ExerciseType",
+    "GradingCriterion",
+    "StructuredGradingInstruction",
     "app"
 ]

athena/athena/schemas/__init__.py

@@ -10,3 +10,4 @@
 from .programming_feedback import ProgrammingFeedback
 from .programming_exercise import ProgrammingExercise
 from .programming_submission import ProgrammingSubmission
+from .grading_criterion import GradingCriterion, StructuredGradingInstruction

athena/athena/schemas/feedback.py

@@ -3,19 +3,23 @@
 
 from pydantic import Field
 
-from .grading_criterion import StructuredGradingInstruction
 from .schema import Schema
 
 
 class Feedback(Schema, ABC):
     id: Optional[int] = Field(None, example=1)
-    title: Optional[str] = Field(None, description="The title of the feedback that is shown to the student.", 
-                      example="File src/pe1/MergeSort.java at line 12")
-    description: Optional[str] = Field(None, description="The detailed feedback description that is shown to the student.",
-                             example="Your solution is correct.")
-    credits: float = Field(0.0, description="The number of points that the student received for this feedback.",
+    title: Optional[str] = Field(None,
+                                 description="The title of the feedback that is shown to the student.",
+                                 example="File src/pe1/MergeSort.java at line 12")
+    description: Optional[str] = Field(None,
+                                       description="The detailed feedback description that is shown to the student.",
+                                       example="Your solution is correct.")
+    credits: float = Field(0.0,
+                           description="The number of points that the student received for this feedback.",
                            example=1.0)
-    structured_grading_instruction_id: Optional[int] = Field(None, description="The id of the structured grading instruction that this feedback belongs to.", example=1)
+    structured_grading_instruction_id: Optional[int] = Field(None,
+                                                             description="The id of the structured grading instruction that this feedback belongs to.",
+                                                             example=1)
 
     meta: dict = Field({}, example={})
 

module_programming_llm/module_programming_llm/generate_suggestions_by_file.py

@@ -29,6 +29,9 @@ class FeedbackModel(BaseModel):
     line_start: Optional[int] = Field(description="Referenced line number start, or empty if unreferenced")
     line_end: Optional[int] = Field(description="Referenced line number end, or empty if unreferenced")
     credits: float = Field(0.0, description="Number of points received/deducted")
+    grading_instruction_id: Optional[int] = Field(
+        description="ID of the grading instruction that was used to generate this feedback, or empty if no grading instruction was used"
+    )
 
     class Config:
         title = "Feedback"
@@ -239,6 +242,7 @@ async def generate_suggestions_by_file(exercise: Exercise, submission: Submissio
                 line_start=feedback.line_start,
                 line_end=feedback.line_end,
                 credits=feedback.credits,
+                structured_grading_instruction_id=feedback.grading_instruction_id,
                 meta={}
             ))
 

module_programming_llm/module_programming_llm/helpers/utils.py

@@ -8,6 +8,7 @@
 from git.repo import Repo
 from langchain.document_loaders import GitLoader
 
+from athena import GradingCriterion
 
 def load_files_from_repo(repo: Repo, file_filter: Optional[Callable[[str], bool]] = None) -> Dict[str, str]:
     return {
@@ -24,6 +25,40 @@ def merge_repos_by_filepath(*repos: Repo, file_filter: Optional[Callable[[str],
         yield (file, [doc.get(file) for doc in docs])
 
 
+def format_grading_instructions(grading_instructions: Optional[str], grading_criteria: Optional[List[GradingCriterion]]) -> Optional[str]:
+    """Formats grading instructions and the grading criteria with nested structured grading instructions into a single string.
+
+    Args:
+        grading_instructions (Optional[str]): Grading instructions
+        grading_criteria (Optional[List[GradingCriterion]]): Grading criteria with nested structured grading instructions
+
+    Returns:
+        Optional[str]: Formatted grading instructions or None if no grading instructions or grading criteria are provided
+    """
+
+    if not grading_instructions and not grading_criteria:
+        return None
+
+    result = ""
+    if grading_instructions:
+        result += grading_instructions + "\n\n"
+
+    if grading_criteria:
+        for grading_criterion in grading_criteria:
+            result += f'Criterion > "{(grading_criterion.title or "Unnamed criterion")}":\n'
+            for grading_instruction in grading_criterion.structured_grading_instructions:
+                result += f'  - grading_instruction_id={grading_instruction.id} > "{grading_instruction.feedback}": ('
+                if grading_instruction.usage_count > 0:
+                    result += f'can be used {grading_instruction.usage_count} times in total'
+                else:
+                    result += "can be used unlimited times"
+                result += f', gives {grading_instruction.credits} credits for "{grading_instruction.grading_scale}" grading scale, '
+                result += f'usage description: "{grading_instruction.instruction_description}")\n'
+            result += "\n"
+
+    return result.strip()
+
+
 def add_line_numbers(content: str) -> str:
     lines = content.splitlines()
     line_number_max_length = len(str(len(lines)))

module_programming_llm/module_programming_llm/split_grading_instructions_by_file.py

@@ -14,7 +14,7 @@
     num_tokens_from_prompt, 
     predict_and_parse
 )
-from module_programming_llm.helpers.utils import get_diff
+from module_programming_llm.helpers.utils import format_grading_instructions, get_diff
 
 
 class FileGradingInstruction(BaseModel):
@@ -47,9 +47,11 @@ async def split_grading_instructions_by_file(
         Optional[SplitGradingInstructions]: Split grading instructions, None if it is too short or too long
     """
 
+    grading_instructions = format_grading_instructions(exercise.grading_instructions, exercise.grading_criteria)
+
     # Return None if the grading instructions are too short
-    if (exercise.grading_instructions is None 
-            or num_tokens_from_string(exercise.grading_instructions) <= config.split_grading_instructions_by_file_prompt.tokens_before_split):
+    if (grading_instructions is None 
+            or num_tokens_from_string(grading_instructions) <= config.split_grading_instructions_by_file_prompt.tokens_before_split):
         return None
 
     # Return None if the grading instructions are not in the prompt
@@ -84,7 +86,7 @@ async def split_grading_instructions_by_file(
     )
 
     prompt_input = {
-        "grading_instructions": exercise.grading_instructions, 
+        "grading_instructions": grading_instructions,
         "changed_files_from_template_to_solution": ", ".join(changed_files_from_template_to_solution),
         "changed_files_from_template_to_submission": ", ".join(changed_files_from_template_to_submission)
     }

module_text_llm/module_text_llm/generate_suggestions.py

@@ -12,14 +12,17 @@
     num_tokens_from_prompt,
     predict_and_parse
 )
-from module_text_llm.helpers.utils import add_sentence_numbers, get_index_range_from_line_range
+from module_text_llm.helpers.utils import add_sentence_numbers, get_index_range_from_line_range, format_grading_instructions
 
 class FeedbackModel(BaseModel):
     title: str = Field(description="Very short title, i.e. feedback category", example="Logic Error")
     description: str = Field(description="Feedback description")
     line_start: Optional[int] = Field(description="Referenced line number start, or empty if unreferenced")
     line_end: Optional[int] = Field(description="Referenced line number end, or empty if unreferenced")
     credits: float = Field(0.0, description="Number of points received/deducted")
+    grading_instruction_id: Optional[int] = Field(
+        description="ID of the grading instruction that was used to generate this feedback, or empty if no grading instruction was used"
+    )
 
     class Config:
         title = "Feedback"
@@ -40,7 +43,7 @@ async def generate_suggestions(exercise: Exercise, submission: Submission, confi
     prompt_input = {
         "max_points": exercise.max_points,
         "bonus_points": exercise.bonus_points,
-        "grading_instructions": exercise.grading_instructions,
+        "grading_instructions": format_grading_instructions(exercise.grading_instructions, exercise.grading_criteria),
         "problem_statement": exercise.problem_statement or "No problem statement.",
         "example_solution": exercise.example_solution,
         "submission": add_sentence_numbers(submission.text)
@@ -102,6 +105,7 @@ async def generate_suggestions(exercise: Exercise, submission: Submission, confi
             index_start=index_start,
             index_end=index_end,
             credits=feedback.credits,
+            structured_grading_instruction_id=feedback.grading_instruction_id,
             meta={}
         ))
 

module_text_llm/module_text_llm/helpers/utils.py

@@ -3,6 +3,8 @@
 import tiktoken
 from nltk.tokenize import sent_tokenize
 
+from athena import GradingCriterion
+
 # This is correct for gpt-4 and chat gpt3.5 but might be different for other models
 def num_tokens_from_string(string: str) -> int:
     """Returns the number of tokens in a text string."""
@@ -11,6 +13,40 @@ def num_tokens_from_string(string: str) -> int:
     return num_tokens
 
 
+def format_grading_instructions(grading_instructions: Optional[str], grading_criteria: Optional[List[GradingCriterion]]) -> Optional[str]:
+    """Formats grading instructions and the grading criteria with nested structured grading instructions into a single string.
+
+    Args:
+        grading_instructions (Optional[str]): Grading instructions
+        grading_criteria (Optional[List[GradingCriterion]]): Grading criteria with nested structured grading instructions
+
+    Returns:
+        Optional[str]: Formatted grading instructions or None if no grading instructions or grading criteria are provided
+    """
+
+    if not grading_instructions and not grading_criteria:
+        return None
+
+    result = ""
+    if grading_instructions:
+        result += grading_instructions + "\n\n"
+
+    if grading_criteria:
+        for grading_criterion in grading_criteria:
+            result += f'Criterion > "{(grading_criterion.title or "Unnamed criterion")}":\n'
+            for grading_instruction in grading_criterion.structured_grading_instructions:
+                result += f'  - grading_instruction_id={grading_instruction.id} > "{grading_instruction.feedback}": ('
+                if grading_instruction.usage_count > 0:
+                    result += f'can be used {grading_instruction.usage_count} times in total'
+                else:
+                    result += "can be used unlimited times"
+                result += f', gives {grading_instruction.credits} credits for "{grading_instruction.grading_scale}" grading scale, '
+                result += f'usage description: "{grading_instruction.instruction_description}")\n'
+            result += "\n"
+
+    return result.strip()
+
+
 def add_sentence_numbers(content: str) -> str:
     sentences = sent_tokenize(content)
     sentences = [line for sentence in sentences for line in sentence.split("\n")]