script/score_by_quota.py

import json
import re
from tqdm import tqdm

data = json.load(open("/root/test/weitiao/data_processing_hsichen/data_process_bq/data/merged_rm_dpo.json"))

def score_response(response_value):
    # Check for incomplete quote nesting (e.g., unmatched * or quotes)
    if response_value.count('*') % 2 != 0 or response_value.count('"') % 2 != 0:
        return 0
    
    # Count narrative (text within *...*) and dialogue (other text)
    narrative_count = len(re.findall(r'\*[^*]+\*', response_value))
    # Remove narrative parts to check for dialogue
    dialogue_text = re.sub(r'\*[^*]+\*', '', response_value).strip()
    dialogue_count = 1 if dialogue_text else 0
    
    # Grading rules
    if narrative_count == 0 and dialogue_count == 1:
        return 2  # Only dialogue
    elif narrative_count >= 1 and dialogue_count == 0:
        return 1  # Only narrative
    elif narrative_count == 1 and dialogue_count == 1:
        return 3  # One narrative + one dialogue
    elif (narrative_count > 1 and dialogue_count >= 1) or (narrative_count >= 1 and dialogue_count > 1):
        return 4  # Multiple narrative/dialogue
    return 0  # Default for incomplete or malformed

# Process each sample
for sample in tqdm(data,total=len(data)):
    chosen_value = sample['chosen']['value']
    rejected_value = sample['rejected']['value']
    
    # Assign scores
    sample['chosen_quota_score'] = score_response(chosen_value)
    sample['rejected_quota_score'] = score_response(rejected_value)

# Save to JSON file
output_file = '/root/test/weitiao/data_processing_hsichen/data_process_bq/data/merged_rm_dpo_scored.json'
with open(output_file, 'w', encoding='utf-8') as f:
    json.dump(data, f, indent=2)

# Print the scored sample for verification