pubmed25_debug.py

# coding=utf-8
# Copyright 2020 The HuggingFace Datasets Authors and the current dataset script contributor.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""MEDLINE/PubMed data - Modified for full abstract text extraction."""


import copy
import gzip
import xml.etree.ElementTree as ET # Using standard ElementTree

import datasets
import random

logger = datasets.logging.get_logger(__name__)


_CITATION = """\
Courtesy of the U.S. National Library of Medicine.
"""

_DESCRIPTION = """\
NLM produces a baseline set of MEDLINE/PubMed citation records in XML format for download on an annual basis.
The annual baseline is released in December of each year. Each day, NLM produces update files that include
new, revised and deleted citations. See our documentation page for more information.
This version is modified to extract the full text from structured abstracts.
"""

_HOMEPAGE = "https://www.nlm.nih.gov/databases/download/pubmed_medline.html"

_LICENSE = "" # Assuming standard NLM terms apply, check source for specifics

# Parameters
total_files = 1274
num_bins = 50
total_urls = 20

# Compute bin size
bin_size = total_files // num_bins

# Sample one random file from each bin
selected_indices = []
for b in range(num_bins):
    start = b * bin_size + 1
    end = min((b + 1) * bin_size + 1, total_files + 1)
    if start < end:
        selected_indices.append(random.randint(start, end - 1))

# Ensure we only keep 20 URLs total (in case rounding leads to more)
selected_indices = sorted(random.sample(selected_indices, total_urls))

# Create URLs
_URLs = [f"https://ftp.ncbi.nlm.nih.gov/pubmed/baseline/pubmed25n{i:04d}.xml.gz" for i in selected_indices]

# Copyright Ferry Boender, released under the MIT license.
# Modified by @Narsil to handle more oddities
def deepupdate(target, src):
    """Deep update target dict with src
    For each k,v in src: if k doesn't exist in target, it is deep copied from
    src to target. Otherwise, if v is a list, target[k] is extended with
    src[k]. If v is a set, target[k] is updated with v, If v is a dict,
    recursively deep-update it.
    """
    for k, v in src.items():
        # Handling type mismatches observed in original script
        if k in target:
            target_type = type(target[k])
            v_type = type(v)

            # Allow string to int conversion for specific known fields if needed
            if target_type == int and v_type == str:
                try:
                    v_int = int(v)
                    v = v_int
                    v_type = int
                except (ValueError, TypeError):
                    logger.warning(f"Field '{k}': Could not convert string '{v}' to expected type {target_type}. Skipping update.")
                    continue # Skip this key if conversion fails

            # If types still don't match after potential conversion, log and skip
            if target_type != v_type and not (isinstance(target[k], list) and isinstance(v, list)) \
               and not (isinstance(target[k], dict) and isinstance(v, dict)) \
               and not (isinstance(target[k], set) and isinstance(v, set)):
                 # Check if target is list and source is not (common issue)
                if isinstance(target[k], list) and not isinstance(v, list):
                     logger.warning(f"Field '{k}': Trying to update a list with a non-list ({v_type}). Wrapping source value in a list.")
                     v = [v] # Attempt to wrap the source value in a list
                # Check if target is dict and source is not
                elif isinstance(target[k], dict) and not isinstance(v, dict):
                     logger.warning(f"Field '{k}': Trying to update a dict with a non-dict ({v_type}). Skipping update.")
                     continue
                # Other mismatches
                else:
                     logger.warning(f"Field '{k}': Type mismatch. Target is {target_type}, Source is {v_type}. Skipping update.")
                     continue


        # Recursive update logic based on type
        if isinstance(v, list):
            if k not in target:
                target[k] = copy.deepcopy(v)
            elif isinstance(target[k], list):
                target[k].extend(v)
            # Handle cases where target is not a list but should be (based on schema usually)
            # This part might be less needed if schema is strictly enforced earlier
            else:
                 logger.warning(f"Field '{k}': Trying to extend a non-list ({type(target[k])}) with a list. Replacing value.")
                 target[k] = copy.deepcopy(v)

        elif isinstance(v, dict):
            if k not in target:
                target[k] = copy.deepcopy(v)
            elif isinstance(target[k], dict):
                deepupdate(target[k], v) # Recurse
            else:
                 logger.warning(f"Field '{k}': Trying to update a non-dict ({type(target[k])}) with a dict. Replacing value.")
                 target[k] = copy.deepcopy(v)

        elif isinstance(v, set):
            if k not in target:
                target[k] = v.copy()
            elif isinstance(target[k], set):
                target[k].update(v.copy())
            else:
                logger.warning(f"Field '{k}': Trying to update a non-set ({type(target[k])}) with a set. Replacing value.")
                target[k] = v.copy()

        # Handle primitive types
        else:
            # Prevent overwriting structured types with primitives if key exists
            if k in target and isinstance(target[k], (list, tuple, dict, set)):
                logger.warning(f"Field '{k}': Trying to overwrite a structured type ({type(target[k])}) with a primitive ({type(v)}). Skipping update.")
                continue
            target[k] = copy.copy(v) # Shallow copy for primitives


def default_date():
    # Using 0 might cause issues with date parsing later. Using None or empty strings might be safer
    # depending on how downstream processing handles missing dates. Sticking to 0 for now.
    return {"Year": 0, "Month": 0, "Day": 0}


def default_inline_article():
    # Default structure expected by the schema
    return {
        # 'Journal': Journal, # Kept commented out as in original
        "Abstract": {"AbstractText": ""}, # Key field for abstract
        "ArticleTitle": "",
        # 'Pagination': {'MedlinePgn': ""}, # Kept commented out
        "AuthorList": {"Author": []}, # Expects a list of Author dicts
        "Language": "",
        "GrantList": {
            "Grant": [], # Expects a list of Grant dicts
        },
        "PublicationTypeList": {"PublicationType": []}, # Expects a list of strings
    }


def default_article():
    # Top-level default structure for a PubmedArticle
    return {
        "MedlineCitation": {
            "PMID": 0, # Needs to be populated
            "DateCompleted": default_date(),
            "NumberOfReferences": 0,
            "DateRevised": default_date(),
            "Article": default_inline_article(), # Embed the article structure
            "MedlineJournalInfo": {"Country": ""},
            "ChemicalList": {"Chemical": []},
            "CitationSubset": "", # Often a list, handle potential type mismatch
            "MeshHeadingList": {"MeshHeading": []},
        },
        "PubmedData": {
            "ArticleIdList": [], # Schema expects Sequence({'ArticleId': Sequence(Value('string'))})
            "PublicationStatus": "",
            "History": {"PubMedPubDate": []},
            "ReferenceList": [], # Modified by update_citation
        },
    }


class Pubmed(datasets.GeneratorBasedBuilder):
    """Pubmed citations records - Modified for full abstract text"""

    # Use a version reflecting the modification, e.g., 5.0.1 or 6.0.0
    # Check the actual NLM baseline year you are targeting
    BUILDER_CONFIGS = [
        datasets.BuilderConfig(name="2025", description="Subset of 2025 annual record with full abstract parsing", version=datasets.Version("6.0.0")),
    ]

    # Keys identified from the schema (_info method) to guide parsing
    SIMPLE_KEYS = set() # Will be populated by fill_keys_from_features
    LIST_KEYS = set()   # Will be populated by fill_keys_from_features
    IGNORE_KEYS = set() # Tracks keys explicitly ignored during parsing

    def fill_keys_from_features(self, features):
        """Recursively populates SIMPLE_KEYS and LIST_KEYS based on the dataset features."""
        if isinstance(features, dict):
            for key, value in features.items():
                if isinstance(value, datasets.Sequence):
                    self.LIST_KEYS.add(key)
                    # Recurse into the sequence's feature type
                    self.fill_keys_from_features(value.feature)
                elif isinstance(value, datasets.Value):
                    # Simple value types (string, int32, etc.)
                    self.SIMPLE_KEYS.add(key)
                elif isinstance(value, dict): # Handle nested dictionaries in features
                    self.SIMPLE_KEYS.add(key) # The key itself points to a dict structure
                    self.fill_keys_from_features(value) # Recurse into the nested dict
                # Add handling for other potential feature types if necessary
        elif isinstance(features, datasets.Sequence): # Handle top-level Sequence feature
             self.fill_keys_from_features(features.feature)

    def get_full_abstract_text(self, abstract_element):
        """
        Extracts and concatenates text from all AbstractText elements
        within the given Abstract element. Handles structured abstracts and inline markup.
        Uses ElementTree's itertext for robustness.
        """
        if abstract_element is None:
            return ""

        # Find all AbstractText child elements directly under Abstract
        # Using XPath './/AbstractText' would find them anywhere below,
        # but './AbstractText' finds only direct children. Let's use direct children first.
        abstract_text_elements = abstract_element.findall('./AbstractText')

        full_text_parts = []
        if not abstract_text_elements:
            # Fallback 1: No direct AbstractText children, maybe text is directly in Abstract? (Unlikely for PubMed)
            # Fallback 2: Or maybe AbstractText tags are nested deeper unexpectedly. Try finding them anywhere.
            abstract_text_elements = abstract_element.findall('.//AbstractText')
            if not abstract_text_elements:
                 # Fallback 3: Use itertext on the Abstract element itself if no AbstractText found
                 all_text = [text.strip() for text in abstract_element.itertext() if text and text.strip()]
                 if all_text:
                      logger.debug(f"Found abstract text directly within <Abstract> or nested tags (no <AbstractText>): {' '.join(all_text)}")
                      return "\n".join(all_text)
                 else:
                      return "" # No text found at all

        # Process found AbstractText elements
        for text_element in abstract_text_elements:
            # Get label attribute if present
            label = text_element.get('Label')
            nlm_category = text_element.get('NlmCategory') # Capture NlmCategory if needed

            # Use itertext() to get *all* text within the element,
            # effectively stripping tags like <i>, <b> but keeping their content.
            # We join fragments within one AbstractText element with spaces,
            # and fragments from different AbstractText elements with newlines later.
            text_content = " ".join(text.strip() for text in text_element.itertext() if text and text.strip())

            # Construct the output string for this part
            current_part = ""
            if label:
                current_part += f"{label}: "
            # Optionally include NlmCategory:
            # if nlm_category:
            #     current_part += f"({nlm_category}) "
            current_part += text_content

            if current_part: # Avoid adding empty parts
                full_text_parts.append(current_part)

        # Join the parts from different AbstractText elements with newlines
        return "\n".join(full_text_parts)


    def xml_to_dictionnary(self, parentElement):
        """
        Recursively converts an XML element and its children into a dictionary,
        guided by SIMPLE_KEYS and LIST_KEYS derived from the dataset schema.
        Includes specific handling for Abstract and ArticleTitle.
        """
        data = {}

        # --- Handling for specific tags ---
        # Handle ArticleTitle: Preserve inner XML tags as a string
        if parentElement.tag == "ArticleTitle":
             # Extract inner XML/HTML as string. itertext() gets only text.
             # To keep tags: concatenate text before first child + string representation of children + tail text
             inner_xml_parts = [ET.tostring(e, encoding='unicode', method='xml') for e in parentElement]
             full_content = (parentElement.text or "").strip() + "".join(inner_xml_parts) + (parentElement.tail or "").strip()
             # Clean up potential extra whitespace between elements
             full_content = ' '.join(full_content.split())
             # Return directly as {Tag: Value} structure expected by caller
             return {parentElement.tag: full_content.strip()}

        # --- Process Children ---
        children = list(parentElement)
        for child in children:
            key = child.tag
            value = None

            # --- Specific Handling for Abstract Tag ---
            if key == "Abstract":
                full_abstract = self.get_full_abstract_text(child)
                # Structure according to schema: {Abstract: {AbstractText: "..."}}
                value = {"AbstractText": full_abstract}
            # Skip AbstractText if found outside Abstract (already handled by get_full_abstract_text)
            elif key == "AbstractText":
                # logger.warning(f"Skipping <{key}> found outside <Abstract> context.")
                continue # Handled within the Abstract block
            # --- End Specific Handling ---
            else:
                # --- Generic Handling for other tags ---
                # Ensure child.text is not None before stripping
                child_text = child.text if child.text is not None else ""

                if len(child) == 0: # Element has no children, just text content
                    value = child_text.strip()
                else: # Element has children, recurse
                    recursive_result = self.xml_to_dictionnary(child)
                    # recursive_result is {'ChildTag': {...}} or {'ChildTag': 'text'}
                    if isinstance(recursive_result, dict) and key in recursive_result:
                         value = recursive_result[key] # Extract the inner value/dict
                    else:
                         # This case might happen if recursion failed or returned unexpected structure
                         logger.warning(f"Unexpected recursive result for <{key}>: {recursive_result}. Using direct text if available.")
                         value = child_text.strip() # Fallback to text

            # --- Add parsed value to data dictionary ---
            if value is None or value == "": # Skip empty/None values unless schema requires empty string
                # Check schema if empty string is valid/expected for this key?
                # For simplicity, skipping None/empty string for now.
                 continue

            if key in data: # Key already exists
                if isinstance(data[key], list):
                    data[key].append(value) # Append to existing list
                else:
                    data[key] = [data[key], value] # Convert to list
            elif key in self.LIST_KEYS:
                # Key expects a list according to schema
                if isinstance(value, list):
                    data[key] = value # Value is already a list (e.g. multiple identical children)
                else:
                    data[key] = [value] # Wrap single value in a list
            elif key in self.SIMPLE_KEYS:
                data[key] = value # Store as simple value
            elif key in self.IGNORE_KEYS:
                continue # Explicitly ignore
            else:
                # Key not in schema or explicitly ignored yet
                # Heuristic: If it appears multiple times, treat as list? Risky.
                # Safest is to log and ignore, or add to IGNORE_KEYS.
                if key not in ["PublicationStatus", "CitationSubset"]: # Allow known schema deviations
                     logger.info(f"Ignoring unexpected key '{key}' found under <{parentElement.tag}>. Content: {value}. Add to Features or IGNORE_KEYS if needed.")
                     self.IGNORE_KEYS.add(key) # Ignore future occurrences in this run
                else: # Handle known deviations that might appear as single strings but schema expects sequence
                     if key in self.LIST_KEYS: # Check if it SHOULD be a list
                          data[key] = [value]
                     else:
                          data[key] = value # Store as simple if not expected list


        # --- Filling Defaults for missing optional elements (as per original script) ---
        if parentElement.tag == "MeshHeading":
            if "QualifierName" not in data: data["QualifierName"] = "" # Expects string
        elif parentElement.tag == "Author":
            if "LastName" not in data: data["LastName"] = ""
            if "ForeName" not in data: data["ForeName"] = ""
            if "Initials" not in data: data["Initials"] = ""
            if "CollectiveName" not in data: data["CollectiveName"] = ""
        elif parentElement.tag == "JournalIssue":
            if "Volume" not in data: data["Volume"] = ""
            if "Issue" not in data: data["Issue"] = ""
            # PubDate handling might be needed if it's inconsistent
        elif parentElement.tag == "Grant":
            if "GrantID" not in data: data["GrantID"] = ""
            # Might need defaults for Agency, Country if they are optional but in schema


        # Return structure expected by recursion {TagName: data_dict}
        # Return empty dict if no children were parsed successfully?
        if data or parentElement.text: # Include if it has data OR direct text (for simple elements)
             # If it only had text content (no children processed), return that directly
             if not data and parentElement.text and parentElement.text.strip():
                  return {parentElement.tag: parentElement.text.strip()}
             else:
                # If it's a structure with attributes, add them
                if parentElement.attrib:
                     # Careful not to overwrite parsed children with same key as attribute
                     attr_dict = {f"@{k}": v for k, v in parentElement.attrib.items() if k not in data}
                     data.update(attr_dict)
                return {parentElement.tag: data}
        else:
             # Return the tag with an empty dict to indicate it was present but empty/unparsed
             return {parentElement.tag: {}}


    def _info(self):
        """Defines the dataset schema."""
        # Define reusable structures
        Date = {
            "Year": datasets.Value("int32"),
            "Month": datasets.Value("int32"),
            "Day": datasets.Value("int32"),
        }

        MeshHeading = {"DescriptorName": datasets.Value("string"), "QualifierName": datasets.Value("string")}

        MedlineJournalInfo = {
            "Country": datasets.Value("string"),
            # Original script commented these out due to inconsistency, keeping them commented
            # 'MedlineTA': datasets.Value('string'),
            # 'NlmUniqueID': datasets.Value('string'),
            # 'ISSNLinking': datasets.Value('string'),
        }
        Chemical = {
            "RegistryNumber": datasets.Value("string"),
            "NameOfSubstance": datasets.Value("string"),
        }

        Author = {
            # Attributes like ValidYN, EqualContrib can be added if needed: "@ValidYN": datasets.Value("string")
            "LastName": datasets.Value("string"),
            "ForeName": datasets.Value("string"),
            "Initials": datasets.Value("string"),
            "CollectiveName": datasets.Value("string"), # Handle cases where Author is CollectiveName
            # AffiliationInfo might be needed
        }
        Grant = {
            # GrantID might be optional in data but required by schema? Check data.
            "GrantID": datasets.Value("string"), # Default added in parser if missing
            "Agency": datasets.Value("string"),
            "Country": datasets.Value("string"),
        }

        # Define the main Article structure, including the corrected Abstract
        Article = {
            # 'Journal': Journal, # Kept commented out
            "Abstract": {"AbstractText": datasets.Value("string")}, # Expects a single string now
            "ArticleTitle": datasets.Value("string"), # Allows HTML tags based on parser
            # 'Pagination': {'MedlinePgn': datasets.Value('string')}, # Kept commented out
            "AuthorList": {"Author": datasets.Sequence(Author)}, # Sequence of Author dicts
            "Language": datasets.Value("string"), # Usually single, but XML allows multiple, use Sequence? Check data. Assuming single for now.
            "GrantList": {
                # GrantList might be optional, Grant inside might be optional/list
                "Grant": datasets.Sequence(Grant), # Sequence of Grant dicts
            },
            "PublicationTypeList": {"PublicationType": datasets.Sequence(datasets.Value("string"))}, # Sequence of strings
        }

        # Define the top-level features including MedlineCitation and PubmedData
        features = datasets.Features(
            {
                "MedlineCitation": {
                    "PMID": datasets.Value("int32"), # Primary key
                    "DateCompleted": Date,
                    "NumberOfReferences": datasets.Value("int32"), # Often missing, default to 0
                    "DateRevised": Date,
                    "Article": Article, # Embed the article structure
                    "MedlineJournalInfo": MedlineJournalInfo,
                    "ChemicalList": {"Chemical": datasets.Sequence(Chemical)},
                    # CitationSubset can be single or multiple - using string for simplicity, may need Sequence
                    "CitationSubset": datasets.Value("string"), # Or datasets.Sequence(datasets.Value("string"))? Check data.
                    "MeshHeadingList": {
                        "MeshHeading": datasets.Sequence(MeshHeading),
                    },
                    # Other potential fields: KeywordList, GeneSymbolList, CommentsCorrectionsList etc.
                },
                "PubmedData": {
                    # ArticleIdList structure is complex: list of dicts, where 'ArticleId' is list of strings
                    "ArticleIdList": datasets.Sequence({
                        "ArticleId": datasets.Sequence(datasets.Value("string")),
                         # Potential attribute: "@IdType": datasets.Value("string")
                        }),
                    "PublicationStatus": datasets.Value("string"),
                    "History": {"PubMedPubDate": datasets.Sequence(Date)}, # Sequence of Date dicts
                    # ReferenceList is modified by update_citation to be simpler
                    "ReferenceList": datasets.Sequence({
                         "Citation": datasets.Value("string"),
                         "CitationId": datasets.Value("int32"), # Assuming PMID is used as CitationId
                         # Potential other fields from original reference structure if needed
                        }),
                },
            }
        )

        # Populate SIMPLE_KEYS and LIST_KEYS based on the final features
        self.fill_keys_from_features(features)

        return datasets.DatasetInfo(
            description=_DESCRIPTION,
            features=features,
            homepage=_HOMEPAGE,
            license=_LICENSE,
            citation=_CITATION,
        )

    def _split_generators(self, dl_manager):
        """Downloads data and defines splits."""
        # Use dl_manager.download_and_extract for Gzip files
        # Using download() and manual gzip opening allows iterparse directly on the stream potentially
        # but download_and_extract is standard practice if disk space is okay.
        # Let's stick to download() and manual gzip opening to facilitate iterparse later.
        dl_dir = dl_manager.download(_URLs) # Returns list/dict of paths to downloaded .xml.gz files
        return [
            datasets.SplitGenerator(
                name=datasets.Split.TRAIN, # Assuming all data goes to train split
                gen_kwargs={"filenames": dl_dir if isinstance(dl_dir, list) else list(dl_dir.values())},
            ),
        ]

    def update_citation(self, article):
        """
        Flattens the complex ReferenceList structure into a simpler list of {Citation, CitationId}.
        Modifies the 'article' dictionary in-place.
        Needs careful error handling for potentially missing keys.
        """
        citations = []
        try:
            # Navigate the potentially complex structure safely
            reference_list_container = article.get("PubmedData", {}).get("ReferenceList")
            if not reference_list_container:
                article["PubmedData"]["ReferenceList"] = [] # Ensure key exists even if empty
                return

            # reference_list_container is likely a list of {'Reference': [...]} dicts
            for ref_container in reference_list_container:
                references = ref_container.get("Reference")
                if not references: continue

                # references is likely a list of actual reference dicts
                for ref in references:
                    citation_text = ref.get("Citation")
                    if not citation_text: continue

                    citation_id = None
                    # Find the PMID in the ArticleIdList within the reference
                    article_id_list_container = ref.get("ArticleIdList")
                    if article_id_list_container:
                         # This is likely a list of {'ArticleId': [...]}
                         for id_container in article_id_list_container:
                              article_ids = id_container.get("ArticleId")
                              if article_ids:
                                   # article_ids is a list of strings, look for integer ones (PMIDs)
                                   for art_id in article_ids:
                                        try:
                                            # Check if it's a digit string and potentially a PMID
                                            if art_id.isdigit():
                                                 citation_id = int(art_id)
                                                 # Optional: Check IdType attribute if available
                                                 # id_type = id_container.get("@IdType")
                                                 # if id_type == "pubmed": # Found explicit PMID
                                                 break # Use the first valid integer ID found
                                        except (ValueError, TypeError):
                                             continue # Skip non-integer IDs
                                   if citation_id: break # Stop looking once an ID is found for this reference
                    # If a valid CitationId (PMID) was found, add the pair
                    if citation_id is not None:
                         citations.append({"Citation": citation_text, "CitationId": citation_id})
                    # else: logger.debug(f"Reference found without parsable PMID: {citation_text[:50]}...")


            # Replace the old ReferenceList structure in PubmedData
            if "PubmedData" in article:
                 article["PubmedData"]["ReferenceList"] = citations
            else:
                 # Handle case where PubmedData might be missing entirely? Unlikely if default used.
                 article["PubmedData"] = {"ReferenceList": citations}


        except Exception as e:
            logger.error(f"Error during citation update for article: {e}")
            # Ensure the key exists even if update fails, matching schema
            if "PubmedData" not in article:
                 article["PubmedData"] = {}
            article["PubmedData"]["ReferenceList"] = []


    def _generate_examples(self, filenames):
        """Yields examples parsing XML files using iterparse for memory efficiency, skipping duplicate PMIDs."""
        # id_ = 0 # Simple counter if needed for logging, but not used as key anymore
        yielded_pmids = set() # Keep track of PMIDs we've already yielded

        for filename in filenames:
            logger.info(f"Processing file: {filename}")
            try:
                with gzip.open(filename, "rb") as f: # Read as bytes for ET
                    context = ET.iterparse(f, events=("end",))
                    event, root = next(context) # Get root iterator

                    for event, elem in context:
                        if event == "end" and elem.tag == "PubmedArticle":
                            article_dict_wrapper = None
                            pmid = "UNKNOWN_PMID" # Default for logging if extraction fails early
                            try:
                                article_dict_wrapper = self.xml_to_dictionnary(elem)

                                if not article_dict_wrapper or 'PubmedArticle' not in article_dict_wrapper:
                                    logger.warning(f"Parser returned empty or invalid structure for a PubmedArticle element in {filename}")
                                    elem.clear()
                                    continue

                                article = article_dict_wrapper.get('PubmedArticle')
                                if not article or not isinstance(article, dict):
                                    logger.warning(f"Parsed empty or invalid article data from element in {filename}")
                                    elem.clear()
                                    continue

                                # --- Extract PMID early for duplicate check ---
                                pmid_val = article.get("MedlineCitation", {}).get("PMID", 0)
                                try:
                                    pmid = int(pmid_val)
                                    if pmid <= 0: raise ValueError("PMID must be positive")
                                except (ValueError, TypeError):
                                    logger.warning(f"Skipping article due to invalid or missing PMID: '{pmid_val}' in {filename}")
                                    elem.clear()
                                    continue # Skip this article if PMID invalid

                                # --- !!! DUPLICATE CHECK !!! ---
                                if pmid in yielded_pmids:
                                    logger.warning(f"Skipping duplicate PMID {pmid} found in {filename}.")
                                    elem.clear()
                                    continue # Skip this duplicate entry
                                # --- End DUPLICATE CHECK ---

                                # --- If not duplicate, proceed with processing ---
                                self.update_citation(article)
                                new_article = default_article()
                                deepupdate(new_article, article)

                                # --- Final validation before yield (PMID check redundant but safe) ---
                                final_pmid_check = new_article.get("MedlineCitation", {}).get("PMID", 0)
                                if final_pmid_check != pmid:
                                     logger.error(f"PMID mismatch after processing! Expected {pmid}, got {final_pmid_check}. Skipping article.")
                                     elem.clear()
                                     continue

                                # Validate against schema
                                encoded_example = self.info.features.encode_example(new_article)

                                # Yield pmid as key and the validated dictionary
                                yield pmid, new_article # Use actual PMID as the example key
                                yielded_pmids.add(pmid) # Add to set *after* successful yield

                            except Exception as e:
                                logger.error(f"Failed to process article PMID {pmid} in {filename}: {e}", exc_info=False) # exc_info=True for full traceback
                                # Logging data causing error can be helpful but verbose:
                                # if 'new_article' in locals(): logger.debug(f"Problematic data: {new_article}")

                            finally:
                                elem.clear() # Clear the element in all cases (success, skip, error)

                    if root is not None:
                        root.clear()

            except ET.ParseError as e:
                logger.error(f"XML ParseError in file {filename}: {e}")
                continue
            except gzip.BadGzipFile:
                logger.error(f"Bad Gzip File error for {filename}. It might be corrupted or incomplete.")
                continue
            except FileNotFoundError:
                 logger.error(f"File not found: {filename}")
                 continue
            except Exception as e:
                logger.error(f"An unexpected error occurred processing file {filename}: {e}", exc_info=True)
                continue