wisent 0.7.701__py3-none-any.whl → 0.7.1045__py3-none-any.whl

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

@@ -6,7 +6,7 @@ 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__ = ["BrowseCompExtractor", "SealExtractor", "FinSearchCompExtractor"]
+__all__ = ["BrowseCompExtractor"]
 
 log = setup_logger(__name__)
 
@@ -44,6 +44,8 @@ class BrowseCompExtractor(HuggingFaceBenchmarkExtractor):
         """
         Build contrastive pairs from BrowseComp examples.
 
+        Uses Tevatron/browsecomp-plus dataset from HuggingFace.
+
         Args:
             limit: Optional maximum number of pairs to produce.
 
@@ -51,74 +53,47 @@ class BrowseCompExtractor(HuggingFaceBenchmarkExtractor):
             A list of ContrastivePair objects.
         """
         max_items = self._normalize_limit(limit)
+        pairs: list[ContrastivePair] = []
 
-        # BrowseComp is not publicly available on HuggingFace
-        # Create synthetic examples based on the benchmark structure
-        docs = self._create_synthetic_examples(max_items or 100)
+        try:
+            docs = self.load_dataset(
+                dataset_name="Tevatron/browsecomp-plus",
+                split="test",
+                limit=max_items,
+            )
+            log.info(f"Loaded {len(docs)} examples from browsecomp-plus")
 
-        pairs: list[ContrastivePair] = []
+            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
 
-        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
+        except Exception as e:
+            log.error(f"Failed to load browsecomp-plus: {e}")
+            return []
 
         if not pairs:
             log.warning("No valid BrowseComp pairs extracted")
 
         return pairs
 
-    def _create_synthetic_examples(self, count: int) -> list[dict[str, Any]]:
-        """Create synthetic BrowseComp examples."""
-        examples = [
-            {
-                "query": "Find the current market cap of Apple Inc.",
-                "correct_answer": "Apple Inc.'s market capitalization can be found on financial websites like Yahoo Finance or Google Finance. As of the most recent data, it is approximately $3 trillion, though this fluctuates with stock price.",
-                "incorrect_answer": "Apple's market cap is $500 billion. I found this on a random website that might not be up to date.",
-                "task_type": "information_retrieval",
-            },
-            {
-                "query": "What are the top 3 most visited websites in the world?",
-                "correct_answer": "According to recent web traffic data from sources like SimilarWeb or Alexa, the top 3 most visited websites are: 1) Google.com, 2) YouTube.com, 3) Facebook.com. These rankings can vary slightly depending on the source and time period.",
-                "incorrect_answer": "The most visited websites are MySpace, Yahoo, and AOL based on what I remember.",
-                "task_type": "fact_finding",
-            },
-            {
-                "query": "Find the official documentation for Python's asyncio library.",
-                "correct_answer": "The official Python asyncio documentation is available at docs.python.org/3/library/asyncio.html. It covers the asyncio module for writing concurrent code using the async/await syntax, including sections on coroutines, tasks, streams, and synchronization primitives.",
-                "incorrect_answer": "I think there's some asyncio documentation somewhere on the internet. You can probably find it by searching.",
-                "task_type": "documentation_search",
-            },
-            {
-                "query": "What is the current weather in Tokyo, Japan?",
-                "correct_answer": "To find current weather in Tokyo, I would check weather.com, accuweather.com, or the Japan Meteorological Agency website (jma.go.jp). These provide real-time weather data including temperature, humidity, and forecasts for Tokyo.",
-                "incorrect_answer": "Tokyo weather is always around 20°C year-round because Japan has a mild climate.",
-                "task_type": "real_time_information",
-            },
-            {
-                "query": "Find research papers about transformer architecture in machine learning.",
-                "correct_answer": "The foundational paper is 'Attention Is All You Need' by Vaswani et al. (2017), available on arXiv (arxiv.org/abs/1706.03762). For more recent research, Google Scholar, arXiv, and Semantic Scholar provide comprehensive collections of transformer-related papers including BERT, GPT, and their variants.",
-                "incorrect_answer": "There are some papers about transformers. They use attention which is like focusing on things.",
-                "task_type": "academic_search",
-            },
-        ]
-
-        result = []
-        for i in range(count):
-            example = examples[i % len(examples)].copy()
-            result.append(example)
-
-        return result
-
     def _extract_pair_from_doc(self, doc: dict[str, Any]) -> ContrastivePair | None:
-        """Convert a single doc into a ContrastivePair."""
+        """Convert a single doc into a ContrastivePair.
+        
+        browsecomp-plus schema:
+        - query_id: str
+        - query: str (the question)
+        - answer: str (ground truth answer)
+        - evidence_docs: list (documents with evidence)
+        - gold_docs: list (gold standard documents)
+        - negative_docs: list (distractor documents)
+        """
         try:
             query = doc.get("query", "").strip()
-            correct_answer = doc.get("correct_answer", "").strip()
-            incorrect_answer = doc.get("incorrect_answer", "").strip()
-            task_type = doc.get("task_type", "general")
+            correct_answer = doc.get("answer", "").strip()
+            query_id = doc.get("query_id", "")
 
             if not query or not correct_answer:
                 return None
@@ -131,10 +106,13 @@ Please search the web and provide accurate, up-to-date information. Include:
 - Relevant details and context
 - Any caveats about data freshness"""
 
+            # Create incorrect answer (opposite or unrelated)
+            incorrect_answer = f"I could not find relevant information about this query."
+
             metadata = {
                 "label": "browsecomp",
-                "source": "browsecomp",
-                "task_type": task_type,
+                "source": "Tevatron/browsecomp-plus",
+                "query_id": query_id,
                 "language": self.language,
                 "is_web_browsing_benchmark": True,
             }
@@ -152,293 +130,5 @@ Please search the web and provide accurate, up-to-date information. Include:
 
 
 
-class SealExtractor(HuggingFaceBenchmarkExtractor):
-    """
-    Extractor for Seal-0 - agentic search evaluation benchmark.
-
-    Seal evaluates LLMs' ability to perform complex multi-step search tasks
-    that require planning, tool use, and information synthesis.
-
-    For agentic search evaluation:
-    - Positive (correct) = Successful multi-step search with correct synthesis
-    - Negative (incorrect) = Failed search or incorrect information gathering
-    """
-
-    # Evaluator that should be used for this benchmark
-    evaluator_name = "agentic_search"
-
-    def extract_contrastive_pairs(
-        self,
-        limit: int | None = None,
-    ) -> list[ContrastivePair]:
-        """
-        Build contrastive pairs from Seal examples.
-
-        Args:
-            limit: Optional maximum number of pairs to produce.
-
-        Returns:
-            A list of ContrastivePair objects.
-        """
-        max_items = self._normalize_limit(limit)
-
-        # Create synthetic examples based on agentic search patterns
-        docs = self._create_synthetic_examples(max_items or 100)
-
-        pairs: list[ContrastivePair] = []
-
-        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 Seal pairs extracted")
-
-        return pairs
-
-    def _create_synthetic_examples(self, count: int) -> list[dict[str, Any]]:
-        """Create synthetic Seal-style examples."""
-        examples = [
-            {
-                "task": "Compare the revenue growth of the top 3 cloud providers over the past 5 years.",
-                "correct_response": """To compare cloud provider revenue growth, I need to:
-1. Search for AWS, Azure, and Google Cloud revenue data
-2. Find historical data from 2019-2024
-3. Calculate growth rates and compare
-
-Based on financial reports and industry analysis:
-- AWS: Grew from ~$35B (2019) to ~$90B (2024), ~157% growth
-- Azure: Grew from ~$17B (2019) to ~$65B (2024), ~282% growth
-- Google Cloud: Grew from ~$8B (2019) to ~$36B (2024), ~350% growth
-
-Key insight: While AWS remains the market leader in absolute revenue, Azure and Google Cloud have shown faster percentage growth rates.""",
-                "incorrect_response": "The cloud providers are all doing well. They make a lot of money from cloud services.",
-                "complexity": "multi_step",
-            },
-            {
-                "task": "Find and summarize the latest developments in quantum computing from the past month.",
-                "correct_response": """I'll search for recent quantum computing news from multiple sources:
-
-Recent developments (November 2024):
-1. IBM announced their 1000+ qubit processor milestone
-2. Google published new research on quantum error correction
-3. IonQ demonstrated improved gate fidelities in trapped-ion systems
-4. Microsoft Azure Quantum expanded their cloud platform capabilities
-
-Sources consulted: Nature, IEEE Spectrum, company press releases, arXiv preprints.
-
-The trend shows continued progress in error correction and qubit count, moving closer to practical quantum advantage for specific applications.""",
-                "incorrect_response": "Quantum computers are getting faster. They use qubits which are like regular bits but quantum.",
-                "complexity": "multi_source",
-            },
-        ]
-
-        result = []
-        for i in range(count):
-            example = examples[i % len(examples)].copy()
-            result.append(example)
-
-        return result
-
-    def _extract_pair_from_doc(self, doc: dict[str, Any]) -> ContrastivePair | None:
-        """Convert a single doc into a ContrastivePair."""
-        try:
-            task = doc.get("task", "").strip()
-            correct = doc.get("correct_response", "").strip()
-            incorrect = doc.get("incorrect_response", "").strip()
-            complexity = doc.get("complexity", "standard")
-
-            if not task or not correct:
-                return None
-
-            task_prompt = f"""Agentic Search Task: {task}
-
-You have access to web search capabilities. Please:
-1. Plan your search strategy
-2. Execute the necessary searches
-3. Synthesize the information into a coherent response
-4. Cite your sources where applicable"""
-
-            metadata = {
-                "label": "seal_0",
-                "source": "seal_0",
-                "complexity": complexity,
-                "is_agentic_search_benchmark": True,
-            }
-
-            return self._build_pair(
-                question=task_prompt,
-                correct=correct,
-                incorrect=incorrect,
-                metadata=metadata,
-            )
-
-        except Exception as exc:
-            log.error(f"Error extracting pair from doc: {exc}", exc_info=True)
-            return None
-
-
-
-class FinSearchCompExtractor(HuggingFaceBenchmarkExtractor):
-    """
-    Extractor for FinSearchComp - financial search agent benchmark.
-
-    FinSearchComp evaluates LLMs' ability to find and analyze financial
-    information, including stock data, financial reports, market analysis,
-    and regulatory filings.
-
-    For financial search evaluation:
-    - Positive (correct) = Accurate financial data with proper sourcing
-    - Negative (incorrect) = Inaccurate data or unsourced claims
-    """
-
-    # Evaluator that should be used for this benchmark
-    evaluator_name = "financial_search"
-
-    def extract_contrastive_pairs(
-        self,
-        limit: int | None = None,
-    ) -> list[ContrastivePair]:
-        """
-        Build contrastive pairs from FinSearchComp examples.
-
-        Args:
-            limit: Optional maximum number of pairs to produce.
-
-        Returns:
-            A list of ContrastivePair objects.
-        """
-        max_items = self._normalize_limit(limit)
-
-        # Create synthetic financial search examples
-        docs = self._create_synthetic_examples(max_items or 100)
-
-        pairs: list[ContrastivePair] = []
-
-        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 FinSearchComp pairs extracted")
-
-        return pairs
-
-    def _create_synthetic_examples(self, count: int) -> list[dict[str, Any]]:
-        """Create synthetic FinSearchComp examples."""
-        examples = [
-            {
-                "query": "What is NVIDIA's P/E ratio and how does it compare to the semiconductor industry average?",
-                "correct_answer": """Based on financial data sources (Yahoo Finance, Bloomberg):
-
-NVIDIA's current P/E ratio: Approximately 65-70x (trailing twelve months)
-Semiconductor industry average P/E: Approximately 25-30x
-
-Analysis: NVIDIA trades at a significant premium to the industry average, reflecting:
-1. Strong growth expectations from AI/datacenter demand
-2. Market leadership in GPU technology
-3. High revenue growth rates (>100% YoY in recent quarters)
-
-Note: P/E ratios fluctuate with stock price and should be verified with real-time data from financial terminals.""",
-                "incorrect_answer": "NVIDIA has a P/E ratio which is a number that shows something about the stock price.",
-                "category": "valuation",
-            },
-            {
-                "query": "Find the key metrics from Tesla's latest quarterly earnings report.",
-                "correct_answer": """Tesla Q3 2024 Earnings Highlights (Source: Tesla Investor Relations, SEC 10-Q):
-
-Revenue: $25.18 billion (+8% YoY)
-Automotive Revenue: $20.02 billion
-Energy & Services Revenue: $5.16 billion
-
-Profitability:
-- Operating Margin: 10.8%
-- Net Income: $2.17 billion
-- EPS: $0.62
-
-Vehicle Deliveries: 462,890 units
-- Model 3/Y: 439,975
-- Other models: 22,915
-
-Key Highlights:
-- Energy storage deployments reached record 6.9 GWh
-- Cybertruck production ramping
-- FSD revenue recognition increasing
-
-Source: Tesla Q3 2024 Update, October 2024""",
-                "incorrect_answer": "Tesla made some money last quarter. They sell cars and batteries.",
-                "category": "earnings",
-            },
-            {
-                "query": "What are the current interest rate expectations for the Federal Reserve?",
-                "correct_answer": """Federal Reserve Interest Rate Outlook (Sources: CME FedWatch, Bloomberg):
-
-Current Federal Funds Rate: 4.50-4.75%
-
-Market Expectations (as of late 2024):
-- December 2024: 75% probability of 25bp cut
-- January 2025: 60% probability of another 25bp cut
-- End of 2025: Terminal rate expected around 3.25-3.50%
-
-Key Factors Driving Expectations:
-1. Inflation trending toward 2% target
-2. Labor market showing signs of cooling
-3. Fed's stated data-dependent approach
-4. Recent Fedspeak suggesting gradual easing path
-
-Note: Interest rate expectations change with economic data releases. Verify with real-time Fed Funds futures.""",
-                "incorrect_answer": "The Fed sets interest rates. They might change them sometime.",
-                "category": "macro",
-            },
-        ]
-
-        result = []
-        for i in range(count):
-            example = examples[i % len(examples)].copy()
-            result.append(example)
-
-        return result
-
-    def _extract_pair_from_doc(self, doc: dict[str, Any]) -> ContrastivePair | None:
-        """Convert a single doc into a ContrastivePair."""
-        try:
-            query = doc.get("query", "").strip()
-            correct = doc.get("correct_answer", "").strip()
-            incorrect = doc.get("incorrect_answer", "").strip()
-            category = doc.get("category", "general")
-
-            if not query or not correct:
-                return None
-
-            task_prompt = f"""Financial Search Task: {query}
 
-Please search for financial data and provide:
-- Specific numbers and metrics where applicable
-- Sources for your data
-- Context and analysis
-- Any caveats about data freshness"""
-
-            metadata = {
-                "label": "finsearchcomp",
-                "source": "finsearchcomp",
-                "category": category,
-                "is_financial_search_benchmark": True,
-            }
-
-            return self._build_pair(
-                question=task_prompt,
-                correct=correct,
-                incorrect=incorrect,
-                metadata=metadata,
-            )
-
-        except Exception as exc:
-            log.error(f"Error extracting pair from doc: {exc}", exc_info=True)
-            return None