GotPsi / src /core /base_classes.py

instNoeticSciences

claude opus 4.6

Flatten history: re-sanitized (128-bit hash) GotPsi public dataset

9deebf2 5 months ago

22 kB

"""
Base classes for the GotPsi data cleaning pipeline.
Defines abstract base classes that establish the interface
for cleaners, processors, and exporters.
"""

from abc import ABC, abstractmethod
from pathlib import Path
from typing import Any, Dict, List, Optional
import warnings
import pandas as pd
from collections import defaultdict

warnings.filterwarnings('ignore', message='.*empty or all-NA entries.*')

from .config import Config
from .errata_logger import ErrataLogger
from .file_discovery_logger import FileDiscoveryLogger


class BaseCleaner(ABC):
    """
    Abstract base class for data cleaning modules.
    Cleaners handle specific aspects of data cleaning such as
    encoding normalization, delimiter standardization, etc.
    """

    def __init__(self, config: Config, errata_logger: Optional[ErrataLogger] = None):
        """
        Initialize cleaner with configuration.
        Args:
            config: Configuration object
            errata_logger: Optional errata logger for error tracking
        """
        self.config = config
        self.errata_logger = errata_logger

    @abstractmethod
    def clean(self, data: Any, **kwargs) -> Any:
        """
        Clean the input data.
        Args:
            data: Input data to clean (format varies by cleaner)
            **kwargs: Additional parameters
        Returns:
            Cleaned data
        """
        pass

    def log_error(
        self,
        error_type: str,
        message: str,
        **kwargs
    ) -> None:
        """
        Log an error to the errata logger if available.
        Args:
            error_type: Type of error
            message: Error message
            **kwargs: Additional context (file_path, row_number, etc.)
        """
        if self.errata_logger:
            self.errata_logger.log_error(error_type, message, **kwargs)


class BaseProcessor(ABC):
    """
    Abstract base class for dataset-specific processors.
    Processors handle the complete data cleaning workflow for
    a specific dataset type (e.g., card tests, RV tests).
    """

    def __init__(self, config: Config, dataset_name: str):
        """
        Initialize processor for a dataset.
        Args:
            config: Configuration object
            dataset_name: Name of the dataset to process
        """
        self.config = config
        self.dataset_name = dataset_name
        self.dataset_config = config.dataset_config(dataset_name)

        # Initialize errata logger
        errata_dir = config.logs_dir / 'errata'
        self.errata_logger = ErrataLogger(errata_dir, dataset_name)

        # Initialize file discovery logger
        file_discovery_dir = config.logs_dir / 'file_discovery'
        self.file_discovery_logger = FileDiscoveryLogger(file_discovery_dir, dataset_name)

    def process(self, max_files: Optional[int] = None, chunk_size: int = 100) -> pd.DataFrame:
        """
        Process all files with parallel cleaning and chunked concatenation.
        Override this in subclasses that need custom post-processing (e.g., users).
        Args:
            max_files: Optional limit on number of files to process
            chunk_size: Number of files to process before concatenating (default: 100)
        Returns:
            Combined DataFrame
        """
        import gc
        from .exceptions import ProcessorError

        files = self.get_file_list()
        if max_files:
            files = files[:max_files]

        print(f"Processing {len(files)} {self.dataset_name} files...")

        # Phase 1: Parallel batch cleaning
        use_delim = getattr(self, 'BATCH_DELIMITER_CLEAN', True)
        cleaned_texts = self._batch_clean_files(files, use_delimiter_cleaner=use_delim)
        print(f"Successfully cleaned {len(cleaned_texts)}/{len(files)} files")

        # Phase 2: Sequential process_file with pre-cleaned text
        intermediate_chunks = []
        current_chunk = []

        try:
            from tqdm import tqdm
            iterator = tqdm(files, desc=f"Processing {self.dataset_name} files")
        except ImportError:
            iterator = files

        for file_path in iterator:
            pre_cleaned = cleaned_texts.get(file_path)
            if pre_cleaned is None:
                continue  # file failed cleaning, already logged

            df = self.process_file(file_path, pre_cleaned_text=pre_cleaned)

            if df is not None and not df.empty:
                current_chunk.append(df)

            if len(current_chunk) >= chunk_size:
                chunk_df = pd.concat(current_chunk, ignore_index=True)
                del current_chunk
                current_chunk = []
                intermediate_chunks.append(chunk_df)
                gc.collect()

        if current_chunk:
            chunk_df = pd.concat(current_chunk, ignore_index=True)
            del current_chunk
            intermediate_chunks.append(chunk_df)
            gc.collect()

        if not intermediate_chunks:
            raise ProcessorError("No valid data processed")

        print(f"Combining {len(intermediate_chunks)} intermediate chunks...")

        # Progressive pair-wise concatenation to minimize peak memory
        while len(intermediate_chunks) > 1:
            pair_results = []
            for i in range(0, len(intermediate_chunks), 2):
                if i + 1 < len(intermediate_chunks):
                    pair_df = pd.concat(
                        [intermediate_chunks[i], intermediate_chunks[i + 1]],
                        ignore_index=True
                    )
                    pair_results.append(pair_df)
                else:
                    pair_results.append(intermediate_chunks[i])
            intermediate_chunks = pair_results
            gc.collect()

        combined = intermediate_chunks[0]
        del intermediate_chunks
        gc.collect()

        # Post-processing hooks
        combined = self._post_process(combined)

        # Close errata logger
        self.errata_logger.close()

        return combined

    def _post_process(self, combined: pd.DataFrame) -> pd.DataFrame:
        """
        Post-processing hook called after all files are combined.
        Default: removes audit columns if audit_mode is disabled.
        Override in subclasses for custom post-processing.
        """
        # Remove audit columns if audit mode is disabled
        audit_mode = self.config.get('processing.audit_mode', False)
        if not audit_mode:
            audit_cols = ['source_file', 'source_row_number']
            cols_to_drop = [col for col in audit_cols if col in combined.columns]
            if cols_to_drop:
                combined = combined.drop(columns=cols_to_drop)
                print(f"Audit mode disabled: Removed columns {cols_to_drop}")
                print()

        return combined

    def pre_validate_csv_lines(
        self,
        text: str,
        file_path: Optional[str] = None,
    ) -> str:
        """Remove lines with inconsistent column counts, logging each drop.
        Determines the modal (most common) comma count, then filters out
        any non-blank line that doesn't match.  This replaces the silent
        on_bad_lines='skip' with an audited pre-filter.
        """
        from collections import Counter

        lines = text.split('\n')
        if not lines:
            return text

        comma_counts = []
        for line in lines:
            stripped = line.strip()
            if stripped:
                comma_counts.append(stripped.count(','))

        if not comma_counts:
            return text

        modal_count = Counter(comma_counts).most_common(1)[0][0]

        kept = []
        dropped = 0
        for lineno, line in enumerate(lines, 1):
            stripped = line.strip()
            if not stripped:
                kept.append(line)
                continue
            if stripped.count(',') == modal_count:
                kept.append(line)
            else:
                dropped += 1
                if dropped <= 20:
                    preview = stripped[:120]
                    self.errata_logger.log_error(
                        'bad_csv_line',
                        f"Line {lineno}: expected {modal_count} commas, got {stripped.count(',')}: {preview!r}",
                        file_path=file_path,
                        line_number=lineno,
                        scope='row',
                    )

        if dropped > 0:
            self.errata_logger.log_error(
                'bad_csv_lines_total',
                f"{dropped} line(s) dropped due to inconsistent column count (expected {modal_count + 1} columns)",
                file_path=file_path,
                scope='file',
            )

        return '\n'.join(kept)

    def to_numeric_logged(
        self,
        series: pd.Series,
        col_name: str,
        file_path: Optional[str] = None,
    ) -> pd.Series:
        """Convert series to numeric, logging any values that get coerced to NaN."""
        coerced = pd.to_numeric(series, errors='coerce')
        bad_mask = coerced.isna() & series.notna() & (series.astype(str).str.strip() != '')
        if bad_mask.any():
            bad_values = series[bad_mask].astype(str).unique()[:10]
            self.errata_logger.log_error(
                'numeric_coercion',
                f"{bad_mask.sum()} non-numeric value(s) coerced to NaN in '{col_name}': {list(bad_values)}",
                file_path=file_path,
                scope='row',
            )
        return coerced

    @abstractmethod
    def process_file(self, file_path: Path, pre_cleaned_text: Optional[str] = None) -> Optional[pd.DataFrame]:
        """
        Process a single file.
        Args:
            file_path: Path to the raw data file
            pre_cleaned_text: Pre-cleaned text (from parallel batch cleaning).
                If None, the processor should clean the file itself.
        Returns:
            DataFrame or None if processing fails
        """
        pass

    # Class attribute: subclasses set this to their glob patterns
    FILE_PATTERNS: List[str] = []

    def get_file_list(self) -> List[Path]:
        """
        Get list of files to process for this dataset.
        Checks config 'file_discovery.use_manifest':
        - False (default): recursive glob using FILE_PATTERNS
        - True: reads from manifest.json in the experiment directory
        Returns:
            List of file paths
        """
        use_manifest = self.config.get('file_discovery.use_manifest', False)

        if use_manifest:
            experiment_dir = self.config.raw_data_dir / self.dataset_name
            return self.discover_files_from_manifest(experiment_dir)
        else:
            files = self.discover_files(self.FILE_PATTERNS)
            if not files:
                from .exceptions import ProcessorError
                raise ProcessorError(
                    f"No {self.dataset_name} files found matching patterns "
                    f"{self.FILE_PATTERNS} in {self.config.raw_data_dir}"
                )
            return files

    def get_dataset_path(self) -> Path:
        """
        Get the full path to the dataset directory.
        Returns:
            Path to dataset directory
        """
        rel_path = self.dataset_config['path']
        return self.config.raw_data_dir / rel_path

    def discover_files(self, patterns: List[str], search_root: Optional[Path] = None) -> List[Path]:
        """
        Recursively discover files matching patterns, with deduplication.
        Searches the data directory recursively for files matching the given patterns.
        When duplicate filenames are found, keeps the larger file.
        Logs all discoveries and deduplication decisions.
        Args:
            patterns: List of filename patterns (e.g., ['users*.dat', 'questions*.dat'])
            search_root: Root directory to search (defaults to config.raw_data_dir)
        Returns:
            List of unique file paths (deduplicated)
        """
        if search_root is None:
            search_root = Path(self.config.raw_data_dir)

        # Log search parameters
        self.file_discovery_logger.set_search_params(search_root, patterns)

        # Discover all matching files
        all_files = []
        for pattern in patterns:
            matching_files = list(search_root.rglob(pattern))
            all_files.extend(matching_files)

        # Log all discovered files
        print(f"File Discovery: Found {len(all_files)} file(s) matching patterns {patterns}")
        for file_path in all_files:
            self.file_discovery_logger.add_discovered_file(file_path)
            print(f"  - {file_path.name}: {file_path.stat().st_size / (1024*1024):.2f} MB")

        # Deduplicate by filename (keep larger file)
        deduplicated_files = self._deduplicate_files(all_files)

        # Log selected files
        print(f"\nAfter deduplication: {len(deduplicated_files)} unique file(s) selected")
        for file_path in deduplicated_files:
            self.file_discovery_logger.add_selected_file(file_path)

        # Finalize and write log
        self.file_discovery_logger.finalize_and_write()
        print(f"File discovery log: {self.file_discovery_logger.get_log_file()}")
        print()

        return deduplicated_files

    def _deduplicate_files(self, files: List[Path]) -> List[Path]:
        """
        Deduplicate files by name, keeping the larger file.
        Args:
            files: List of file paths
        Returns:
            Deduplicated list of file paths
        """
        # Group files by name
        files_by_name = defaultdict(list)
        for file_path in files:
            files_by_name[file_path.name].append(file_path)

        # Select one file per name (largest)
        selected_files = []
        for filename, file_list in files_by_name.items():
            if len(file_list) == 1:
                # No duplicates, just add it
                selected_files.append(file_list[0])
            else:
                # Multiple files with same name, keep largest
                largest_file = max(file_list, key=lambda f: f.stat().st_size)
                selected_files.append(largest_file)

                # Log the duplicate set
                self.file_discovery_logger.add_duplicate_set(
                    files=file_list,
                    selected_file=largest_file,
                    reason=f"Largest file ({largest_file.stat().st_size} bytes)"
                )

                # Print warning about duplicates
                print(f"\n⚠ Found {len(file_list)} copies of '{filename}':")
                for f in file_list:
                    marker = "✓ SELECTED" if f == largest_file else "  skipped"
                    print(f"  {marker}: {f} ({f.stat().st_size / (1024*1024):.2f} MB)")

        return sorted(selected_files)

    def discover_files_from_manifest(self, experiment_dir: Path) -> List[Path]:
        """
        Discover files by reading the experiment directory's manifest.json.
        Args:
            experiment_dir: Path to the experiment directory containing manifest.json
        Returns:
            List of file paths that exist on disk
        Raises:
            ProcessorError: If no manifest-listed files exist on disk
        """
        import json
        import logging
        from .exceptions import ProcessorError

        logger = logging.getLogger(__name__)

        manifest_path = experiment_dir / 'manifest.json'
        if not manifest_path.exists():
            raise ProcessorError(
                f"No manifest.json found in {experiment_dir}. "
                f"Run scripts/migrate_raw_data.py first."
            )

        manifest = json.loads(manifest_path.read_text())
        found = []
        missing = []

        for entry in manifest.get('files', []):
            file_path = experiment_dir / entry['filename']
            if file_path.exists():
                found.append(file_path)
            else:
                missing.append(entry['filename'])
                logger.warning(f"Manifest-listed file not found: {file_path}")

        if missing:
            logger.warning(
                f"{len(missing)}/{len(manifest['files'])} manifest files missing "
                f"in {experiment_dir}"
            )

        if not found:
            raise ProcessorError(
                f"No manifest-listed files found in {experiment_dir}. "
                f"Run scripts/download_data.py to fetch the data."
            )

        return sorted(found)

    def _read_and_clean_text(
        self,
        file_path: Path,
        use_delimiter_cleaner: bool = True
    ) -> str:
        """
        Read a file and clean its encoding and delimiters.
        This is the unit of work for parallel batch cleaning.
        Args:
            file_path: Path to the raw data file
            use_delimiter_cleaner: Whether to apply delimiter standardization
        Returns:
            Cleaned UTF-8 text string
        """
        # Step 1: encoding clean (detect + decode + mojibake fix)
        text = self.encoding_cleaner.clean(file_path)

        # Step 2: optional delimiter standardization
        if use_delimiter_cleaner and hasattr(self, 'delimiter_cleaner'):
            text = self.delimiter_cleaner.clean(text, file_path=str(file_path))

        return text

    def _batch_clean_files(
        self,
        files: List[Path],
        use_delimiter_cleaner: bool = True,
        max_workers: Optional[int] = None
    ) -> Dict[Path, str]:
        """
        Clean all files in parallel using ThreadPoolExecutor.
        Encoding detection (chardet) and file I/O release the GIL,
        making threads effective here.
        Args:
            files: List of file paths to clean
            use_delimiter_cleaner: Whether to apply delimiter standardization
            max_workers: Max threads (defaults to min(8, len(files)))
        Returns:
            Dict mapping file path to cleaned text. Failed files are excluded.
        """
        from concurrent.futures import ThreadPoolExecutor, as_completed

        if not files:
            return {}

        if max_workers is None:
            max_workers = min(8, len(files))

        results: Dict[Path, str] = {}

        def clean_one(file_path: Path) -> tuple:
            try:
                text = self._read_and_clean_text(file_path, use_delimiter_cleaner)
                return (file_path, text, None)
            except Exception as e:
                return (file_path, None, e)

        with ThreadPoolExecutor(max_workers=max_workers) as executor:
            futures = {executor.submit(clean_one, f): f for f in files}

            try:
                from tqdm import tqdm
                iterator = tqdm(
                    as_completed(futures),
                    total=len(futures),
                    desc="Cleaning files"
                )
            except ImportError:
                iterator = as_completed(futures)

            for future in iterator:
                file_path, text, error = future.result()
                if error is not None:
                    self.errata_logger.log_error(
                        'batch_clean_failed',
                        f'Failed to clean file: {error}',
                        file_path=str(file_path),
                        scope='file'
                    )
                else:
                    results[file_path] = text

        return results


class BaseProfiler(ABC):
    """
    Abstract base class for file profilers.
    Profilers analyze files to detect format characteristics
    such as encoding, delimiters, and schema.
    """

    def __init__(self, config: Config):
        """
        Initialize profiler with configuration.
        Args:
            config: Configuration object
        """
        self.config = config

    @abstractmethod
    def profile(self, file_path: str | Path) -> Dict[str, Any]:
        """
        Profile a file to detect its characteristics.
        Args:
            file_path: Path to file to profile
        Returns:
            Dictionary with profile information
        """
        pass


class BaseExporter(ABC):
    """
    Abstract base class for data exporters.
    Exporters handle writing cleaned data to various formats
    (HDF5, Parquet, CSV).
    """

    def __init__(self, config: Config):
        """
        Initialize exporter with configuration.
        Args:
            config: Configuration object
        """
        self.config = config

    @abstractmethod
    def export(self, data: pd.DataFrame, output_path: str | Path, **kwargs) -> None:
        """
        Export data to output file.
        Args:
            data: DataFrame to export
            output_path: Output file path
            **kwargs: Format-specific parameters
        """
        pass


class BaseQualityAssessor(ABC):
    """
    Abstract base class for data quality assessment.
    Quality assessors analyze cleaned data and compute
    quality scores and metrics.
    """

    def __init__(self, config: Config):
        """
        Initialize quality assessor with configuration.
        Args:
            config: Configuration object
        """
        self.config = config

    @abstractmethod
    def assess(self, data: pd.DataFrame) -> Dict[str, Any]:
        """
        Assess data quality and compute metrics.
        Args:
            data: DataFrame to assess
        Returns:
            Dictionary with quality metrics
        """
        pass

    @abstractmethod
    def compute_quality_score(self, metrics: Dict[str, Any]) -> float:
        """
        Compute overall quality score from metrics.
        Args:
            metrics: Quality metrics dictionary
        Returns:
            Quality score (0-10)
        """
        pass

	"""
	Base classes for the GotPsi data cleaning pipeline.

	Defines abstract base classes that establish the interface
	for cleaners, processors, and exporters.
	"""

	from abc import ABC, abstractmethod
	from pathlib import Path
	from typing import Any, Dict, List, Optional
	import warnings
	import pandas as pd
	from collections import defaultdict

	warnings.filterwarnings('ignore', message='.empty or all-NA entries.')

	from .config import Config
	from .errata_logger import ErrataLogger
	from .file_discovery_logger import FileDiscoveryLogger


	class BaseCleaner(ABC):
	"""
	Abstract base class for data cleaning modules.

	Cleaners handle specific aspects of data cleaning such as
	encoding normalization, delimiter standardization, etc.
	"""

	def __init__(self, config: Config, errata_logger: Optional[ErrataLogger] = None):
	"""
	Initialize cleaner with configuration.

	Args:
	config: Configuration object
	errata_logger: Optional errata logger for error tracking
	"""
	self.config = config
	self.errata_logger = errata_logger

	@abstractmethod
	def clean(self, data: Any, **kwargs) -> Any:
	"""
	Clean the input data.

	Args:
	data: Input data to clean (format varies by cleaner)
	**kwargs: Additional parameters

	Returns:
	Cleaned data
	"""
	pass

	def log_error(
	self,
	error_type: str,
	message: str,
	**kwargs
	) -> None:
	"""
	Log an error to the errata logger if available.

	Args:
	error_type: Type of error
	message: Error message
	**kwargs: Additional context (file_path, row_number, etc.)
	"""
	if self.errata_logger:
	self.errata_logger.log_error(error_type, message, **kwargs)


	class BaseProcessor(ABC):
	"""
	Abstract base class for dataset-specific processors.

	Processors handle the complete data cleaning workflow for
	a specific dataset type (e.g., card tests, RV tests).
	"""

	def __init__(self, config: Config, dataset_name: str):
	"""
	Initialize processor for a dataset.

	Args:
	config: Configuration object
	dataset_name: Name of the dataset to process
	"""
	self.config = config
	self.dataset_name = dataset_name
	self.dataset_config = config.dataset_config(dataset_name)

	# Initialize errata logger
	errata_dir = config.logs_dir / 'errata'
	self.errata_logger = ErrataLogger(errata_dir, dataset_name)

	# Initialize file discovery logger
	file_discovery_dir = config.logs_dir / 'file_discovery'
	self.file_discovery_logger = FileDiscoveryLogger(file_discovery_dir, dataset_name)

	def process(self, max_files: Optional[int] = None, chunk_size: int = 100) -> pd.DataFrame:
	"""
	Process all files with parallel cleaning and chunked concatenation.

	Override this in subclasses that need custom post-processing (e.g., users).

	Args:
	max_files: Optional limit on number of files to process
	chunk_size: Number of files to process before concatenating (default: 100)

	Returns:
	Combined DataFrame
	"""
	import gc
	from .exceptions import ProcessorError

	files = self.get_file_list()
	if max_files:
	files = files[:max_files]

	print(f"Processing {len(files)} {self.dataset_name} files...")

	# Phase 1: Parallel batch cleaning
	use_delim = getattr(self, 'BATCH_DELIMITER_CLEAN', True)
	cleaned_texts = self._batch_clean_files(files, use_delimiter_cleaner=use_delim)
	print(f"Successfully cleaned {len(cleaned_texts)}/{len(files)} files")

	# Phase 2: Sequential process_file with pre-cleaned text
	intermediate_chunks = []
	current_chunk = []

	try:
	from tqdm import tqdm
	iterator = tqdm(files, desc=f"Processing {self.dataset_name} files")
	except ImportError:
	iterator = files

	for file_path in iterator:
	pre_cleaned = cleaned_texts.get(file_path)
	if pre_cleaned is None:
	continue # file failed cleaning, already logged

	df = self.process_file(file_path, pre_cleaned_text=pre_cleaned)

	if df is not None and not df.empty:
	current_chunk.append(df)

	if len(current_chunk) >= chunk_size:
	chunk_df = pd.concat(current_chunk, ignore_index=True)
	del current_chunk
	current_chunk = []
	intermediate_chunks.append(chunk_df)
	gc.collect()

	if current_chunk:
	chunk_df = pd.concat(current_chunk, ignore_index=True)
	del current_chunk
	intermediate_chunks.append(chunk_df)
	gc.collect()

	if not intermediate_chunks:
	raise ProcessorError("No valid data processed")

	print(f"Combining {len(intermediate_chunks)} intermediate chunks...")

	# Progressive pair-wise concatenation to minimize peak memory
	while len(intermediate_chunks) > 1:
	pair_results = []
	for i in range(0, len(intermediate_chunks), 2):
	if i + 1 < len(intermediate_chunks):
	pair_df = pd.concat(
	[intermediate_chunks[i], intermediate_chunks[i + 1]],
	ignore_index=True
	)
	pair_results.append(pair_df)
	else:
	pair_results.append(intermediate_chunks[i])
	intermediate_chunks = pair_results
	gc.collect()

	combined = intermediate_chunks[0]
	del intermediate_chunks
	gc.collect()

	# Post-processing hooks
	combined = self._post_process(combined)

	# Close errata logger
	self.errata_logger.close()

	return combined

	def _post_process(self, combined: pd.DataFrame) -> pd.DataFrame:
	"""
	Post-processing hook called after all files are combined.

	Default: removes audit columns if audit_mode is disabled.
	Override in subclasses for custom post-processing.
	"""
	# Remove audit columns if audit mode is disabled
	audit_mode = self.config.get('processing.audit_mode', False)
	if not audit_mode:
	audit_cols = ['source_file', 'source_row_number']
	cols_to_drop = [col for col in audit_cols if col in combined.columns]
	if cols_to_drop:
	combined = combined.drop(columns=cols_to_drop)
	print(f"Audit mode disabled: Removed columns {cols_to_drop}")
	print()

	return combined

	def pre_validate_csv_lines(
	self,
	text: str,
	file_path: Optional[str] = None,
	) -> str:
	"""Remove lines with inconsistent column counts, logging each drop.

	Determines the modal (most common) comma count, then filters out
	any non-blank line that doesn't match. This replaces the silent
	on_bad_lines='skip' with an audited pre-filter.
	"""
	from collections import Counter

	lines = text.split('\n')
	if not lines:
	return text

	comma_counts = []
	for line in lines:
	stripped = line.strip()
	if stripped:
	comma_counts.append(stripped.count(','))

	if not comma_counts:
	return text

	modal_count = Counter(comma_counts).most_common(1)[0][0]

	kept = []
	dropped = 0
	for lineno, line in enumerate(lines, 1):
	stripped = line.strip()
	if not stripped:
	kept.append(line)
	continue
	if stripped.count(',') == modal_count:
	kept.append(line)
	else:
	dropped += 1
	if dropped <= 20:
	preview = stripped[:120]
	self.errata_logger.log_error(
	'bad_csv_line',
	f"Line {lineno}: expected {modal_count} commas, got {stripped.count(',')}: {preview!r}",
	file_path=file_path,
	line_number=lineno,
	scope='row',
	)

	if dropped > 0:
	self.errata_logger.log_error(
	'bad_csv_lines_total',
	f"{dropped} line(s) dropped due to inconsistent column count (expected {modal_count + 1} columns)",
	file_path=file_path,
	scope='file',
	)

	return '\n'.join(kept)

	def to_numeric_logged(
	self,
	series: pd.Series,
	col_name: str,
	file_path: Optional[str] = None,
	) -> pd.Series:
	"""Convert series to numeric, logging any values that get coerced to NaN."""
	coerced = pd.to_numeric(series, errors='coerce')
	bad_mask = coerced.isna() & series.notna() & (series.astype(str).str.strip() != '')
	if bad_mask.any():
	bad_values = series[bad_mask].astype(str).unique()[:10]
	self.errata_logger.log_error(
	'numeric_coercion',
	f"{bad_mask.sum()} non-numeric value(s) coerced to NaN in '{col_name}': {list(bad_values)}",
	file_path=file_path,
	scope='row',
	)
	return coerced

	@abstractmethod
	def process_file(self, file_path: Path, pre_cleaned_text: Optional[str] = None) -> Optional[pd.DataFrame]:
	"""
	Process a single file.

	Args:
	file_path: Path to the raw data file
	pre_cleaned_text: Pre-cleaned text (from parallel batch cleaning).
	If None, the processor should clean the file itself.

	Returns:
	DataFrame or None if processing fails
	"""
	pass

	# Class attribute: subclasses set this to their glob patterns
	FILE_PATTERNS: List[str] = []

	def get_file_list(self) -> List[Path]:
	"""
	Get list of files to process for this dataset.

	Checks config 'file_discovery.use_manifest':
	- False (default): recursive glob using FILE_PATTERNS
	- True: reads from manifest.json in the experiment directory

	Returns:
	List of file paths
	"""
	use_manifest = self.config.get('file_discovery.use_manifest', False)

	if use_manifest:
	experiment_dir = self.config.raw_data_dir / self.dataset_name
	return self.discover_files_from_manifest(experiment_dir)
	else:
	files = self.discover_files(self.FILE_PATTERNS)
	if not files:
	from .exceptions import ProcessorError
	raise ProcessorError(
	f"No {self.dataset_name} files found matching patterns "
	f"{self.FILE_PATTERNS} in {self.config.raw_data_dir}"
	)
	return files

	def get_dataset_path(self) -> Path:
	"""
	Get the full path to the dataset directory.

	Returns:
	Path to dataset directory
	"""
	rel_path = self.dataset_config['path']
	return self.config.raw_data_dir / rel_path

	def discover_files(self, patterns: List[str], search_root: Optional[Path] = None) -> List[Path]:
	"""
	Recursively discover files matching patterns, with deduplication.

	Searches the data directory recursively for files matching the given patterns.
	When duplicate filenames are found, keeps the larger file.
	Logs all discoveries and deduplication decisions.

	Args:
	patterns: List of filename patterns (e.g., ['users.dat', 'questions.dat'])
	search_root: Root directory to search (defaults to config.raw_data_dir)

	Returns:
	List of unique file paths (deduplicated)
	"""
	if search_root is None:
	search_root = Path(self.config.raw_data_dir)

	# Log search parameters
	self.file_discovery_logger.set_search_params(search_root, patterns)

	# Discover all matching files
	all_files = []
	for pattern in patterns:
	matching_files = list(search_root.rglob(pattern))
	all_files.extend(matching_files)

	# Log all discovered files
	print(f"File Discovery: Found {len(all_files)} file(s) matching patterns {patterns}")
	for file_path in all_files:
	self.file_discovery_logger.add_discovered_file(file_path)
	print(f" - {file_path.name}: {file_path.stat().st_size / (1024*1024):.2f} MB")

	# Deduplicate by filename (keep larger file)
	deduplicated_files = self._deduplicate_files(all_files)

	# Log selected files
	print(f"\nAfter deduplication: {len(deduplicated_files)} unique file(s) selected")
	for file_path in deduplicated_files:
	self.file_discovery_logger.add_selected_file(file_path)

	# Finalize and write log
	self.file_discovery_logger.finalize_and_write()
	print(f"File discovery log: {self.file_discovery_logger.get_log_file()}")
	print()

	return deduplicated_files

	def _deduplicate_files(self, files: List[Path]) -> List[Path]:
	"""
	Deduplicate files by name, keeping the larger file.

	Args:
	files: List of file paths

	Returns:
	Deduplicated list of file paths
	"""
	# Group files by name
	files_by_name = defaultdict(list)
	for file_path in files:
	files_by_name[file_path.name].append(file_path)

	# Select one file per name (largest)
	selected_files = []
	for filename, file_list in files_by_name.items():
	if len(file_list) == 1:
	# No duplicates, just add it
	selected_files.append(file_list[0])
	else:
	# Multiple files with same name, keep largest
	largest_file = max(file_list, key=lambda f: f.stat().st_size)
	selected_files.append(largest_file)

	# Log the duplicate set
	self.file_discovery_logger.add_duplicate_set(
	files=file_list,
	selected_file=largest_file,
	reason=f"Largest file ({largest_file.stat().st_size} bytes)"
	)

	# Print warning about duplicates
	print(f"\n⚠ Found {len(file_list)} copies of '{filename}':")
	for f in file_list:
	marker = "✓ SELECTED" if f == largest_file else " skipped"
	print(f" {marker}: {f} ({f.stat().st_size / (1024*1024):.2f} MB)")

	return sorted(selected_files)

	def discover_files_from_manifest(self, experiment_dir: Path) -> List[Path]:
	"""
	Discover files by reading the experiment directory's manifest.json.

	Args:
	experiment_dir: Path to the experiment directory containing manifest.json

	Returns:
	List of file paths that exist on disk

	Raises:
	ProcessorError: If no manifest-listed files exist on disk
	"""
	import json
	import logging
	from .exceptions import ProcessorError

	logger = logging.getLogger(__name__)

	manifest_path = experiment_dir / 'manifest.json'
	if not manifest_path.exists():
	raise ProcessorError(
	f"No manifest.json found in {experiment_dir}. "
	f"Run scripts/migrate_raw_data.py first."
	)

	manifest = json.loads(manifest_path.read_text())
	found = []
	missing = []

	for entry in manifest.get('files', []):
	file_path = experiment_dir / entry['filename']
	if file_path.exists():
	found.append(file_path)
	else:
	missing.append(entry['filename'])
	logger.warning(f"Manifest-listed file not found: {file_path}")

	if missing:
	logger.warning(
	f"{len(missing)}/{len(manifest['files'])} manifest files missing "
	f"in {experiment_dir}"
	)

	if not found:
	raise ProcessorError(
	f"No manifest-listed files found in {experiment_dir}. "
	f"Run scripts/download_data.py to fetch the data."
	)

	return sorted(found)

	def _read_and_clean_text(
	self,
	file_path: Path,
	use_delimiter_cleaner: bool = True
	) -> str:
	"""
	Read a file and clean its encoding and delimiters.

	This is the unit of work for parallel batch cleaning.

	Args:
	file_path: Path to the raw data file
	use_delimiter_cleaner: Whether to apply delimiter standardization

	Returns:
	Cleaned UTF-8 text string
	"""
	# Step 1: encoding clean (detect + decode + mojibake fix)
	text = self.encoding_cleaner.clean(file_path)

	# Step 2: optional delimiter standardization
	if use_delimiter_cleaner and hasattr(self, 'delimiter_cleaner'):
	text = self.delimiter_cleaner.clean(text, file_path=str(file_path))

	return text

	def _batch_clean_files(
	self,
	files: List[Path],
	use_delimiter_cleaner: bool = True,
	max_workers: Optional[int] = None
	) -> Dict[Path, str]:
	"""
	Clean all files in parallel using ThreadPoolExecutor.

	Encoding detection (chardet) and file I/O release the GIL,
	making threads effective here.

	Args:
	files: List of file paths to clean
	use_delimiter_cleaner: Whether to apply delimiter standardization
	max_workers: Max threads (defaults to min(8, len(files)))

	Returns:
	Dict mapping file path to cleaned text. Failed files are excluded.
	"""
	from concurrent.futures import ThreadPoolExecutor, as_completed

	if not files:
	return {}

	if max_workers is None:
	max_workers = min(8, len(files))

	results: Dict[Path, str] = {}

	def clean_one(file_path: Path) -> tuple:
	try:
	text = self._read_and_clean_text(file_path, use_delimiter_cleaner)
	return (file_path, text, None)
	except Exception as e:
	return (file_path, None, e)

	with ThreadPoolExecutor(max_workers=max_workers) as executor:
	futures = {executor.submit(clean_one, f): f for f in files}

	try:
	from tqdm import tqdm
	iterator = tqdm(
	as_completed(futures),
	total=len(futures),
	desc="Cleaning files"
	)
	except ImportError:
	iterator = as_completed(futures)

	for future in iterator:
	file_path, text, error = future.result()
	if error is not None:
	self.errata_logger.log_error(
	'batch_clean_failed',
	f'Failed to clean file: {error}',
	file_path=str(file_path),
	scope='file'
	)
	else:
	results[file_path] = text

	return results


	class BaseProfiler(ABC):
	"""
	Abstract base class for file profilers.

	Profilers analyze files to detect format characteristics
	such as encoding, delimiters, and schema.
	"""

	def __init__(self, config: Config):
	"""
	Initialize profiler with configuration.

	Args:
	config: Configuration object
	"""
	self.config = config

	@abstractmethod
	def profile(self, file_path: str \| Path) -> Dict[str, Any]:
	"""
	Profile a file to detect its characteristics.

	Args:
	file_path: Path to file to profile

	Returns:
	Dictionary with profile information
	"""
	pass


	class BaseExporter(ABC):
	"""
	Abstract base class for data exporters.

	Exporters handle writing cleaned data to various formats
	(HDF5, Parquet, CSV).
	"""

	def __init__(self, config: Config):
	"""
	Initialize exporter with configuration.

	Args:
	config: Configuration object
	"""
	self.config = config

	@abstractmethod
	def export(self, data: pd.DataFrame, output_path: str \| Path, **kwargs) -> None:
	"""
	Export data to output file.

	Args:
	data: DataFrame to export
	output_path: Output file path
	**kwargs: Format-specific parameters
	"""
	pass


	class BaseQualityAssessor(ABC):
	"""
	Abstract base class for data quality assessment.

	Quality assessors analyze cleaned data and compute
	quality scores and metrics.
	"""

	def __init__(self, config: Config):
	"""
	Initialize quality assessor with configuration.

	Args:
	config: Configuration object
	"""
	self.config = config

	@abstractmethod
	def assess(self, data: pd.DataFrame) -> Dict[str, Any]:
	"""
	Assess data quality and compute metrics.

	Args:
	data: DataFrame to assess

	Returns:
	Dictionary with quality metrics
	"""
	pass

	@abstractmethod
	def compute_quality_score(self, metrics: Dict[str, Any]) -> float:
	"""
	Compute overall quality score from metrics.

	Args:
	metrics: Quality metrics dictionary

	Returns:
	Quality score (0-10)
	"""
	pass