wisent 0.7.379__py3-none-any.whl → 0.7.701__py3-none-any.whl

wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/instruct_humaneval.py

@@ -1,180 +0,0 @@
-from __future__ import annotations
-
-from typing import Any
-from wisent.core.cli_logger import setup_logger
-
-from wisent.core.contrastive_pairs.core.pair import ContrastivePair
-from wisent.core.contrastive_pairs.huggingface_pairs.atoms import HuggingFaceBenchmarkExtractor
-
-__all__ = ["InstructHumanevalExtractor"]
-
-log = setup_logger(__name__)
-
-
-class InstructHumanevalExtractor(HuggingFaceBenchmarkExtractor):
-    """
-    Extractor for instruct_humaneval dataset.
-
-    Schema (openai_humaneval):
-        - prompt: str (question/prompt)
-        - canonical_solution: str (answer/solution)
-    """
-
-    def extract_contrastive_pairs(
-        self,
-        limit: int | None = None,
-    ) -> list[ContrastivePair]:
-        """
-        Build contrastive pairs from instruct_humaneval examples.
-
-        Args:
-            limit: Optional maximum number of pairs to produce.
-
-        Returns:
-            A list of ContrastivePair objects.
-        """
-        max_items = self._normalize_limit(limit)
-
-        # Load dataset
-        docs = self.load_dataset(
-            dataset_name="openai_humaneval",
-            split="test",
-            limit=max_items,
-        )
-
-        pairs: list[ContrastivePair] = []
-
-        log.info(f"Extracting contrastive pairs from {len(docs)} instruct_humaneval examples")
-
-        for doc in docs:
-            pair = self._extract_pair_from_doc(doc)
-            if pair is not None:
-                pairs.append(pair)
-                if max_items is not None and len(pairs) >= max_items:
-                    break
-
-        if not pairs:
-            log.warning("No valid instruct_humaneval pairs extracted")
-
-        return pairs
-
-    def _extract_pair_from_doc(self, doc: dict[str, Any]) -> ContrastivePair | None:
-        """
-        Convert a single doc into a ContrastivePair.
-
-        Returns None when required fields are missing or malformed.
-        """
-        try:
-            question = doc.get("prompt", "").strip()
-            answer = doc.get("canonical_solution", "")
-            test_code = doc.get("test", "").strip()
-            entry_point = doc.get("entry_point", "")
-
-            if not question or not answer:
-                log.debug("Skipping: missing question or answer")
-                return None
-
-            # HumanEval canonical_solution is just the function body
-            # We need to combine it with the function signature from the prompt
-            # to create a complete executable function
-
-            # First corrupt the function body BEFORE building the complete function
-            # This ensures we corrupt the actual code, not the docstring
-            correct_body = str(answer).strip()
-            incorrect_body = self._create_incorrect_answer(correct_body)
-
-            # Build complete functions with correct and incorrect bodies
-            correct_answer = self._build_complete_function(question, correct_body)
-            incorrect_answer = self._build_complete_function(question, incorrect_body)
-
-            # Format the question
-            formatted_question = f"Question: {question}\n\nWhat is the answer?"
-
-            metadata = {
-                "label": "instruct_humaneval",
-                "source": "openai_humaneval",
-                "test_code": test_code,  # Include test code for execution
-                "entry_point": entry_point,
-            }
-
-            return self._build_pair(
-                question=formatted_question,
-                correct=correct_answer,
-                incorrect=incorrect_answer,
-                metadata=metadata,
-            )
-
-        except Exception as exc:
-            log.error(f"Error extracting pair from doc: {exc}", exc_info=True)
-            return None
-
-    def _build_complete_function(self, prompt: str, function_body: str) -> str:
-        """Build a complete function by combining prompt signature with function body.
-
-        HumanEval prompt contains:
-        - imports (e.g., from typing import List)
-        - function signature (e.g., def func(args) -> ReturnType:)
-        - docstring
-
-        The canonical_solution contains only the function body (indented code).
-
-        We need to combine them to create an executable function.
-        """
-        import re
-
-        # Extract the function definition line (starts with 'def ')
-        # This includes everything from 'def' up to and including the colon
-        def_pattern = r'(def\s+\w+\([^)]*\)(?:\s*->\s*[^:]+)?:)'
-        def_match = re.search(def_pattern, prompt, re.MULTILINE)
-
-        if not def_match:
-            # Fallback: return body as-is (might fail, but better than crashing)
-            log.warning("Could not extract function signature from prompt")
-            return function_body
-
-        function_signature = def_match.group(1)
-
-        # Extract any imports before the function definition
-        imports_section = prompt[:def_match.start()].strip()
-
-        # Extract docstring if present (text between the function signature and body)
-        after_signature = prompt[def_match.end():].strip()
-        docstring = ""
-        if after_signature.startswith('"""') or after_signature.startswith("'''"):
-            # Find the closing quotes
-            quote_char = '"""' if after_signature.startswith('"""') else "'''"
-            end_quote = after_signature.find(quote_char, len(quote_char))
-            if end_quote != -1:
-                docstring = "    " + after_signature[:end_quote + len(quote_char)]
-
-        # Build the complete function
-        parts = []
-        if imports_section:
-            parts.append(imports_section)
-            parts.append("")  # Blank line after imports
-
-        parts.append(function_signature)
-
-        if docstring:
-            parts.append(docstring)
-
-        # Add the function body
-        # The canonical_solution from HumanEval has its first line at column 0,
-        # but subsequent lines are properly indented. We need to add 4 spaces
-        # to the first line to match the indentation of other lines.
-        if function_body:
-            # Add indentation to the first line if it's not already indented
-            body_lines = function_body.split('\n')
-            if body_lines and body_lines[0] and not body_lines[0].startswith('    '):
-                body_lines[0] = '    ' + body_lines[0]
-            parts.append('\n'.join(body_lines))
-
-        return "\n".join(parts)
-
-    def _create_incorrect_answer(self, correct: str) -> str:
-        """Create an incorrect answer by modifying the correct one."""
-        # For code, corrupt it slightly by adding a comment in the middle
-        if len(correct) > 10:
-            return correct[:len(correct)//2] + "\n           # CORRUPTED" + correct[len(correct)//2:]
-        return f"{correct} # INCORRECT"
-

wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/instructhumaneval.py

@@ -1,129 +0,0 @@
-from __future__ import annotations
-
-from typing import Any
-from wisent.core.cli_logger import setup_logger
-
-from wisent.core.contrastive_pairs.core.pair import ContrastivePair
-from wisent.core.contrastive_pairs.huggingface_pairs.atoms import HuggingFaceBenchmarkExtractor
-
-__all__ = ["InstructHumanEvalExtractor"]
-
-log = setup_logger(__name__)
-
-# Task names this extractor handles
-task_names = ("instruct_humaneval",)
-
-# Evaluator to use
-class InstructHumanEvalExtractor(HuggingFaceBenchmarkExtractor):
-    evaluator_name = "coding"
-    """
-    Extractor for InstructHumanEval coding benchmark.
-
-    InstructHumanEval schema (codeparrot/instructhumaneval):
-        - task_id: str (e.g., "HumanEval/0")
-        - prompt: str (function signature + docstring)
-        - canonical_solution: str (correct implementation)
-        - test: str (unit tests)
-        - entry_point: str (function name)
-        - instruction: str (natural language instruction)
-        - signature: str (function signature)
-        - docstring: str (function docstring)
-        - context: str (function signature without body)
-    """
-
-    def extract_contrastive_pairs(
-        self,
-        limit: int | None = None,
-    ) -> list[ContrastivePair]:
-        """
-        Build contrastive pairs from InstructHumanEval examples.
-
-        For coding tasks, we create pairs where:
-        - Positive: Correct implementation
-        - Negative: Incorrect implementation (placeholder)
-
-        Args:
-            limit: Optional maximum number of pairs to produce.
-
-        Returns:
-            A list of ContrastivePair objects.
-        """
-        max_items = self._normalize_limit(limit)
-
-        # Load InstructHumanEval dataset
-        docs = self.load_dataset(
-            dataset_name="codeparrot/instructhumaneval",
-            split="test",
-            limit=max_items,
-        )
-
-        pairs: list[ContrastivePair] = []
-
-        log.info(f"Extracting contrastive pairs from {len(docs)} InstructHumanEval examples")
-
-        for doc in docs:
-            pair = self._extract_pair_from_doc(doc)
-            if pair is not None:
-                pairs.append(pair)
-                if max_items is not None and len(pairs) >= max_items:
-                    break
-
-        if not pairs:
-            log.warning("No valid InstructHumanEval pairs extracted")
-
-        return pairs
-
-    def _extract_pair_from_doc(self, doc: dict[str, Any]) -> ContrastivePair | None:
-        """
-        Convert a single InstructHumanEval doc into a ContrastivePair.
-
-        Returns None when required fields are missing or malformed.
-        """
-        try:
-            task_id = doc.get("task_id", "")
-            prompt = doc.get("prompt", "").strip()
-            canonical_solution = doc.get("canonical_solution", "").strip()
-            entry_point = doc.get("entry_point", "")
-            test_code = doc.get("test", "").strip()
-            instruction = doc.get("instruction", "").strip()
-
-            if not prompt or not canonical_solution:
-                log.debug(f"Skipping {task_id}: missing prompt or solution")
-                return None
-
-            # Construct the full correct implementation
-            # In InstructHumanEval, prompt contains signature+docstring, canonical_solution is body at column 0
-            # We need to indent the body by 4 spaces to be inside the function
-            lines = canonical_solution.split('\n')
-            indented_lines = ['    ' + line if line and not line[0].isspace() else line for line in lines]
-            indented_solution = '\n'.join(indented_lines)
-            correct_code = prompt + "\n" + indented_solution
-
-            # Create an incorrect implementation (return incorrect value/type)
-            # For coding benchmarks, we create a simple buggy version
-            incorrect_code = prompt + "\n    pass  # Incorrect: empty implementation"
-
-            # Format the question using the instruction field if available
-            if instruction:
-                question = f"{instruction}\n\n{prompt}"
-            else:
-                question = f"Complete the following Python function:\n\n{prompt}"
-
-            metadata = {
-                "label": "instruct_humaneval",
-                "task_id": task_id,
-                "entry_point": entry_point,
-                "test_code": test_code,  # Include test code for execution
-            }
-
-            return self._build_pair(
-                question=question,
-                correct=correct_code,
-                incorrect=incorrect_code,
-                metadata=metadata,
-            )
-
-        except Exception as exc:
-            log.error(f"Error extracting pair from doc: {exc}", exc_info=True)
-            return None
-

wisent/core/contrastive_pairs/huggingface_pairs/hf_task_extractors/mbpp.py

@@ -1,142 +0,0 @@
-from __future__ import annotations
-
-from typing import Any
-from wisent.core.cli_logger import setup_logger
-
-from wisent.core.contrastive_pairs.core.pair import ContrastivePair
-from wisent.core.contrastive_pairs.huggingface_pairs.atoms import HuggingFaceBenchmarkExtractor
-
-__all__ = ["MBPPExtractor"]
-
-log = setup_logger(__name__)
-
-task_names = ("mbpp_plus",)
-
-class MBPPExtractor(HuggingFaceBenchmarkExtractor):
-    """
-    Extractor for MBPP (Mostly Basic Python Problems) dataset.
-
-    Schema (mbpp or google-research/mbpp):
-        - text: str (problem description)
-        - code: str (correct solution code)
-        - test_list: list[str] (test cases)
-    """
-
-
-    evaluator_name = "coding"
-
-    def extract_contrastive_pairs(
-        self,
-        limit: int | None = None,
-    ) -> list[ContrastivePair]:
-        """
-        Build contrastive pairs from MBPP examples.
-
-        Args:
-            limit: Optional maximum number of pairs to produce.
-
-        Returns:
-            A list of ContrastivePair objects.
-        """
-        max_items = self._normalize_limit(limit)
-
-        # Load dataset
-        docs = self.load_dataset(
-            dataset_name="mbpp",
-            split="test",
-            limit=max_items,
-        )
-
-        pairs: list[ContrastivePair] = []
-
-        log.info(f"Extracting contrastive pairs from {len(docs)} MBPP examples")
-
-        for doc in docs:
-            pair = self._extract_pair_from_doc(doc)
-            if pair is not None:
-                pairs.append(pair)
-                if max_items is not None and len(pairs) >= max_items:
-                    break
-
-        if not pairs:
-            log.warning("No valid MBPP pairs extracted")
-
-        return pairs
-
-    def _extract_pair_from_doc(self, doc: dict[str, Any]) -> ContrastivePair | None:
-        """
-        Convert a single doc into a ContrastivePair.
-
-        Returns None when required fields are missing or malformed.
-        """
-        try:
-            problem_text = doc.get("text", "").strip()
-            correct_code = doc.get("code", "").strip()
-            test_list = doc.get("test_list", [])
-            test_imports = doc.get("test_imports", [])
-
-            if not problem_text or not correct_code:
-                log.debug("Skipping: missing problem text or code")
-                return None
-
-            # Create incorrect code (add syntax error or logical error)
-            incorrect_code = self._create_incorrect_code(correct_code)
-
-            # Format the prompt
-            formatted_prompt = f"{problem_text}\n\nWrite a Python function to solve this problem."
-
-            # Build test code from test_list
-            # The test_list contains assertion strings like "assert function_name(...) == expected"
-            test_code = ""
-            entry_point = None
-            if test_list:
-                # Extract function name from correct_code
-                import re
-                func_match = re.search(r'def\s+(\w+)\s*\(', correct_code)
-                if func_match:
-                    entry_point = func_match.group(1)
-                
-                # Add imports if provided
-                if test_imports:
-                    test_code += "\n".join(test_imports) + "\n\n"
-
-                # Build test function that uses 'candidate' parameter
-                test_code += "def check(candidate):\n"
-                for test_case in test_list:
-                    # Replace function name with 'candidate' in assertions
-                    if entry_point:
-                        modified_test = test_case.replace(f"{entry_point}(", "candidate(")
-                    else:
-                        modified_test = test_case
-                    test_code += f"    {modified_test}\n"
-
-            metadata = {
-                "label": "mbpp",
-                "source": "mbpp",
-                "test_code": test_code if test_code else None,
-                "entry_point": entry_point,
-            }
-
-            return self._build_pair(
-                question=formatted_prompt,
-                correct=correct_code,
-                incorrect=incorrect_code,
-                metadata=metadata,
-            )
-
-        except Exception as exc:
-            log.error(f"Error extracting pair from doc: {exc}", exc_info=True)
-            return None
-
-    def _create_incorrect_code(self, correct: str) -> str:
-        """Create an incorrect version of the code."""
-        # Add a syntax error by removing closing parenthesis
-        if "(" in correct and ")" in correct:
-            # Find last closing paren and remove it
-            idx = correct.rfind(")")
-            if idx > 0:
-                return correct[:idx] + correct[idx+1:]
-
-        # Fallback: add comment that breaks the code
-        return correct + "\n# Missing return statement"
-

wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/agieval.py

@@ -1,155 +0,0 @@
-from __future__ import annotations
-
-from typing import Any, TYPE_CHECKING
-
-from wisent.core.contrastive_pairs.core.pair import ContrastivePair
-from wisent.core.contrastive_pairs.core.response import NegativeResponse, PositiveResponse
-from wisent.core.contrastive_pairs.lm_eval_pairs.atoms import LMEvalBenchmarkExtractor
-from wisent.core.cli_logger import setup_logger, bind
-
-if TYPE_CHECKING:
-    from lm_eval.api.task import ConfigurableTask
-
-
-__all__ = ["AgievalExtractor"]
-_LOG = setup_logger(__name__)
-
-task_names = ("agieval",)
-
-class AgievalExtractor(LMEvalBenchmarkExtractor):
-    """Extractor for the Agieval benchmark."""
-
-
-    evaluator_name = "exact_match"
-    def extract_contrastive_pairs(
-        self,
-        lm_eval_task_data: ConfigurableTask,
-        limit: int | None = None,
-        preferred_doc: str | None = None,
-    ) -> list[ContrastivePair]:
-        """
-        Build contrastive pairs from Agieval docs.
-
-        Args:
-            lm_eval_task_data: lm-eval task instance for Agieval.
-            limit: Optional maximum number of pairs to produce.
-            preferred_doc: Optional preferred document source.
-
-        Returns:
-            A list of ContrastivePair objects.
-        """
-        log = bind(_LOG, task=getattr(lm_eval_task_data, "NAME", "unknown"))
-
-        max_items = self._normalize_limit(limit)
-        docs = self.load_docs(lm_eval_task_data, max_items, preferred_doc=preferred_doc)
-
-        pairs: list[ContrastivePair] = []
-
-        log.info("Extracting contrastive pairs", extra={"doc_count": len(docs)})
-
-        for doc in docs:
-            pair = self._extract_pair_from_doc(doc)
-            if pair is not None:
-                pairs.append(pair)
-                if max_items is not None and len(pairs) >= max_items:
-                    break
-
-        if not pairs:
-            task_name = getattr(lm_eval_task_data, "NAME", type(lm_eval_task_data).__name__)
-            log.warning("No valid Agieval pairs extracted", extra={"task": task_name})
-
-        return pairs
-
-    def _extract_pair_from_doc(self, doc: dict[str, Any]) -> ContrastivePair | None:
-        """
-        Convert a single Agieval doc into a ContrastivePair, if possible.
-        Returns None when required fields are missing or malformed.
-        """
-        log = bind(_LOG, doc_id=doc.get("id", "unknown"))
-
-        try:
-            # Try multiple possible schema formats
-            question = None
-            choices = None
-            answer_idx = None
-
-            # Format 1: question + choices + answer
-            if "question" in doc and "choices" in doc:
-                question = str(doc.get("question", "")).strip()
-                choices_data = doc.get("choices", {})
-                if isinstance(choices_data, dict):
-                    choices = choices_data.get("text", [])
-                elif isinstance(choices_data, list):
-                    choices = choices_data
-                answer = doc.get("answer", doc.get("answerKey", ""))
-                if isinstance(answer, str) and len(answer) == 1 and answer.isalpha():
-                    answer_idx = ord(answer.upper()) - ord('A')
-                else:
-                    answer_idx = int(answer) if answer else 0
-
-            # Format 2: instruction + option_a/b/c/d + answer (MMMLU style)
-            elif "instruction" in doc and "option_a" in doc:
-                question = str(doc.get("instruction", "")).strip()
-                choices = [
-                    str(doc.get("option_a", "")).strip(),
-                    str(doc.get("option_b", "")).strip(),
-                    str(doc.get("option_c", "")).strip(),
-                    str(doc.get("option_d", "")).strip(),
-                ]
-                choices = [c for c in choices if c]
-                answer = doc.get("answer", "A")
-                answer_idx = ord(str(answer).upper()) - ord('A')
-
-            # Format 3: query/prompt + answer
-            elif "query" in doc or "prompt" in doc:
-                question = str(doc.get("query", doc.get("prompt", ""))).strip()
-                # For open-ended questions, use target as correct answer
-                correct_answer = str(doc.get("target", doc.get("answer", ""))).strip()
-                if correct_answer:
-                    metadata = {"label": "agieval"}
-                    return self._build_pair(
-                        question=f"Question: {question}",
-                        correct=correct_answer,
-                        incorrect="incorrect answer",
-                        metadata=metadata,
-                    )
-                return None
-
-            if not question or not choices or answer_idx is None or not (0 <= answer_idx < len(choices)):
-                log.debug(
-                    "Skipping doc due to missing/invalid fields",
-                    extra={"doc": doc},
-                )
-                return None
-
-            correct = choices[answer_idx]
-            incorrect_idx = (answer_idx + 1) % len(choices)
-            incorrect = choices[incorrect_idx]
-
-            formatted_question = f"Question: {question}\nA. {incorrect}\nB. {correct}"
-
-            metadata = {
-                "label": "agieval",
-            }
-
-            return self._build_pair(
-                question=formatted_question,
-                correct=correct,
-                incorrect=incorrect,
-                metadata=metadata,
-            )
-
-        except Exception as exc:
-            log.error("Error extracting pair from doc", exc_info=exc, extra={"doc": doc})
-            return None
-
-    @staticmethod
-    def _build_pair(
-        question: str,
-        correct: str,
-        incorrect: str,
-        metadata: dict[str, Any] | None = None,
-    ) -> ContrastivePair:
-        positive_response = PositiveResponse(model_response=correct)
-        negative_response = NegativeResponse(model_response=incorrect)
-        return ContrastivePair(prompt=question, positive_response=positive_response, negative_response=negative_response, label=metadata.get("label"))