Datasets:

NbAiLab
/

nynorsk-subcorpora

+import json
+import re
+from typing import List
+from jiwer import wer # Make sure jiwer is installed (pip install jiwer)
+from tqdm import tqdm
+import argparse
+def extract_html_tags(text: str) -> List[str]:
+    """Extracts HTML-like tags from a given text string."""
+    return re.findall(r"</?[^>]+?>", text)
+def tags_to_str(tags: List[str]) -> str:
+    """Converts a list of tag strings into a single space-separated string."""
+    return " ".join(tags)
+def calculate_tag_reward(source_text: str, target_text: str) -> float:
+    """
+    Calculates the tag preservation reward between a source and target text.
+    Reward is 1.0 if Tag WER is 0, and 1.0 - Tag WER otherwise (capped at 0).
+    """
+    source_tags_list = extract_html_tags(source_text)
+    target_tags_list = extract_html_tags(target_text)
+    source_tags_str = tags_to_str(source_tags_list)
+    target_tags_str = tags_to_str(target_tags_list)
+    tag_error_rate = wer(source_tags_str, target_tags_str)
+    # Ensure reward is between 0 and 1
+    reward = 1.0 - min(1.0, max(0.0, tag_error_rate))
+    if reward < 1.0: # If there's any error, print details
+        print(f"\n--- Mismatch Found (Reward: {reward:.4f}, Tag WER: {tag_error_rate:.4f}) ---")
+        print(f"Source Text (snippet): {source_text[:100]}...")
+        print(f"Target Text (snippet): {target_text[:100]}...")
+        print(f"Source Tags (str): '{source_tags_str}'")
+        print(f"Target Tags (str): '{target_tags_str}'")
+        # For more detailed debugging of WER, you can print the lists too:
+        # print(f"Source Tags (list): {source_tags_list}")
+        # print(f"Target Tags (list): {target_tags_list}")
+        print("----------------------------------------------------")
+    return reward
+def main(jsonl_file_path: str):
+    non_one_reward_count = 0
+    total_lines_processed = 0
+    print(f"Processing file: {jsonl_file_path}")
+    with open(jsonl_file_path, 'r', encoding='utf-8') as f:
+        # First count lines for tqdm progress bar
+        num_lines = sum(1 for line in f)
+        f.seek(0) # Reset file pointer to the beginning
+        for line in tqdm(f, total=num_lines, desc="Verifying dataset tags"):
+            try:
+                data = json.loads(line)
+                nb_text = data.get("nb")
+                nn_text = data.get("nn")
+                if nb_text is None or nn_text is None:
+                    print(f"Warning: Missing 'nb' or 'nn' field in line: {line.strip()}")
+                    continue
+                total_lines_processed += 1
+                reward = calculate_tag_reward(nb_text, nn_text)
+                if reward < 1.0:
+                    non_one_reward_count += 1
+            except json.JSONDecodeError:
+                print(f"Warning: Could not decode JSON from line: {line.strip()}")
+            except Exception as e:
+                print(f"An error occurred processing line: {line.strip()} - Error: {e}")
+    print(f"\n--- Verification Complete ---")
+    print(f"Total lines processed: {total_lines_processed}")
+    print(f"Number of lines where tag reward (nb vs nn) was NOT 1.0: {non_one_reward_count}")
+    if total_lines_processed > 0:
+        percentage_mismatch = (non_one_reward_count / total_lines_processed) * 100
+        print(f"Percentage of lines with tag mismatches: {percentage_mismatch:.2f}%")
+    else:
+        print("No lines were processed.")
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="Verify HTML tag consistency between 'nb' and 'nn' fields in a JSONL dataset.")
+    parser.add_argument("jsonl_file", type=str, help="Path to the .jsonl dataset file.")
+    args = parser.parse_args()
+    main(args.jsonl_file)