wisent 0.7.379__py3-none-any.whl → 0.7.901__py3-none-any.whl

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"))

wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/code2text.py

@@ -1,161 +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__ = ["Code2textExtractor"]
-_LOG = setup_logger(__name__)
-
-task_names = (
-    "code2text_go",
-    "code2text_java",
-    "code2text_javascript",
-    "code2text_php",
-    "code2text_python",
-    "code2text_ruby",
-)
-
-class Code2textExtractor(LMEvalBenchmarkExtractor):
-    """Extractor for Code2Text benchmark."""
-
-
-    evaluator_name = "generation"
-    def extract_contrastive_pairs(
-        self,
-        lm_eval_task_data: ConfigurableTask,
-        limit: int | None = None,
-        preferred_doc: str | None = None,
-    ) -> list[ContrastivePair]:
-        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 pairs extracted", extra={"task": task_name})
-
-        return pairs
-
-    def _extract_pair_from_doc(self, doc: dict[str, Any]) -> ContrastivePair | None:
-        log = bind(_LOG, doc_id=doc.get("id", "unknown"))
-
-        try:
-            # Format 1: Code2Text format (code_tokens + docstring_tokens)
-            if "code_tokens" in doc and "docstring_tokens" in doc:
-                code_tokens = doc.get("code_tokens", [])
-                docstring_tokens = doc.get("docstring_tokens", [])
-
-                if not code_tokens or not docstring_tokens:
-                    log.debug("Skipping doc - empty code or docstring tokens", extra={"doc": doc})
-                    return None
-
-                # Join tokens to create code and docstring
-                code = " ".join(code_tokens).replace("\n", " ")
-                code = " ".join(code.strip().split())
-
-                docstring = " ".join(docstring_tokens).replace("\n", " ")
-                docstring = " ".join(docstring.strip().split())
-
-                # Create synthetic negative by shuffling words in the docstring
-                import random
-                words = docstring.split()
-                if len(words) > 3:
-                    shuffled_words = words.copy()
-                    random.shuffle(shuffled_words)
-
-                    # If shuffle resulted in same order, reverse it
-                    if shuffled_words == words:
-                        shuffled_words = list(reversed(words))
-
-                    incorrect_docstring = " ".join(shuffled_words)
-                else:
-                    # For very short docstrings, just use a generic incorrect one
-                    incorrect_docstring = "This is an incorrect docstring."
-
-                prompt = f"Generate documentation for this code:\n\n{code}\n\nDocumentation:"
-
-                metadata = {"label": "code2text"}
-
-                return self._build_pair(
-                    question=prompt,
-                    correct=docstring,
-                    incorrect=incorrect_docstring,
-                    metadata=metadata,
-                )
-
-            # Format 2: Multiple choice format (fallback)
-            # Try multiple format patterns for question
-            question = doc.get("question", doc.get("query", doc.get("input", doc.get("instruction", doc.get("prompt", ""))))).strip()
-
-            # Try multiple format patterns for choices
-            choices = doc.get("choices", doc.get("options", doc.get("answers", [])))
-
-            # Handle option_a/b/c/d format
-            if not choices and "option_a" in doc:
-                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]
-
-            # Try multiple format patterns for answer
-            answer = doc.get("answer", doc.get("label", doc.get("target", None)))
-
-            if isinstance(answer, str) and len(answer) == 1 and answer.isalpha():
-                answer_idx = ord(answer.upper()) - ord('A')
-            elif isinstance(answer, int):
-                answer_idx = answer
-            else:
-                return None
-
-            if not question or not choices or not (0 <= answer_idx < len(choices)):
-                log.debug("Skipping doc due to missing/invalid fields", extra={"doc": doc})
-                return None
-
-            correct = str(choices[answer_idx]).strip()
-            incorrect_idx = (answer_idx + 1) % len(choices)
-            incorrect = str(choices[incorrect_idx]).strip()
-
-            formatted_question = f"Question: {question}\nA. {incorrect}\nB. {correct}"
-            metadata = {"label": "code2text"}
-
-            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"))

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

@@ -1,107 +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__ = ["CodexglueExtractor"]
-_LOG = setup_logger(__name__)
-
-
-class CodexglueExtractor(LMEvalBenchmarkExtractor):
-    """Extractor for Codexglue benchmark."""
-
-    def extract_contrastive_pairs(
-        self,
-        lm_eval_task_data: ConfigurableTask,
-        limit: int | None = None,
-        preferred_doc: str | None = None,
-    ) -> list[ContrastivePair]:
-        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 pairs extracted", extra={"task": task_name})
-
-        return pairs
-
-    def _extract_pair_from_doc(self, doc: dict[str, Any]) -> ContrastivePair | None:
-        log = bind(_LOG, doc_id=doc.get("id", "unknown"))
-
-        try:
-            # Try multiple format patterns for question
-            question = doc.get("question", doc.get("query", doc.get("input", doc.get("instruction", doc.get("prompt", ""))))).strip()
-            
-            # Try multiple format patterns for choices
-            choices = doc.get("choices", doc.get("options", doc.get("answers", [])))
-            
-            # Handle option_a/b/c/d format
-            if not choices and "option_a" in doc:
-                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]
-
-            # Try multiple format patterns for answer
-            answer = doc.get("answer", doc.get("label", doc.get("target", None)))
-
-            if isinstance(answer, str) and len(answer) == 1 and answer.isalpha():
-                answer_idx = ord(answer.upper()) - ord('A')
-            elif isinstance(answer, int):
-                answer_idx = answer
-            else:
-                return None
-
-            if not question or not choices or not (0 <= answer_idx < len(choices)):
-                log.debug("Skipping doc due to missing/invalid fields", extra={"doc": doc})
-                return None
-
-            correct = str(choices[answer_idx]).strip()
-            incorrect_idx = (answer_idx + 1) % len(choices)
-            incorrect = str(choices[incorrect_idx]).strip()
-
-            formatted_question = f"Question: {question}\nA. {incorrect}\nB. {correct}"
-            metadata = {"label": "codexglue"}
-
-            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"))

wisent/core/contrastive_pairs/lm_eval_pairs/lm_task_extractors/livemathbench.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__ = ["LivemathbenchExtractor"]
-_LOG = setup_logger(__name__)
-
-task_names = ("livemathbench", "livemathbench_cnmo_en", "livemathbench_cnmo_zh")
-
-class LivemathbenchExtractor(LMEvalBenchmarkExtractor):
-    """Extractor for Livemathbench benchmark (live math problems)."""
-
-
-    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 Livemathbench docs.
-
-        Args:
-            lm_eval_task_data: lm-eval task instance for Livemathbench.
-            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 Livemathbench pairs extracted", extra={"task": task_name})
-
-        return pairs
-
-    def _extract_pair_from_doc(self, doc: dict[str, Any]) -> ContrastivePair | None:
-        """
-        Convert a single Livemathbench 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
-            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/problem + answer (common for math)
-            elif "query" in doc or "prompt" in doc or "problem" in doc:
-                question = str(doc.get("query", doc.get("prompt", doc.get("problem", "")))).strip()
-                # For open-ended questions, use target as correct answer
-                correct_answer = str(doc.get("target", doc.get("answer", doc.get("solution", "")))).strip()
-                if correct_answer:
-                    metadata = {"label": "livemathbench"}
-                    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": "livemathbench",
-            }
-
-            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"))