lphatnguyen/distillation
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

modify gen_vqa_bank

Browse Source
This commit is contained in:
lphatnguyen
2025-08-07 15:45:55 +00:00
parent 0393ae3ab4
commit 814fbfee03
1 changed files with 6 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
easydistill/mmkd/dev-vqa/gen_vqa_bank.py
View File

@@ -7,7 +7,7 @@ import re
def load_json(filepath): def load_json(filepath):
""" """
Loads a JSON file with robust error handling. Loads a JSON file .
""" """
try: try:
with open(filepath, 'r', encoding='utf-8') as f: with open(filepath, 'r', encoding='utf-8') as f:
@@ -21,7 +21,7 @@ def load_json(filepath):
def read_text_file(filepath): def read_text_file(filepath):
""" """
Loads a simple text file. Loads a prompt from a text file.
""" """
try: try:
with open(filepath, 'r', encoding='utf-8') as f: with open(filepath, 'r', encoding='utf-8') as f:
@@ -32,7 +32,7 @@ def read_text_file(filepath):
def format_items_list(items, language): def format_items_list(items, language):
""" """
Formats a list of item dictionaries into a human-readable string. Formats a list of item dictionaries (services) into a human-readable string.
""" """
if not items: if not items:
return "" return ""
@@ -92,7 +92,7 @@ def get_conversational_answer(field, label_data, answer_bank, language):
return str(value) if value is not None else "" return str(value) if value is not None else ""
# --- Conversations Generation for Label Data --- # --- Conversations Generation for Label Data ---
def generate_field_level_conversations(labels_path, image_root, system_prompt_path, questions_path, answers_path, output_path): def generate_vqa_conversations(labels_path, image_root, system_prompt_path, questions_path, answers_path, output_path):
""" """
Generates multiple conversational VQA pairs for each field in a label file, Generates multiple conversational VQA pairs for each field in a label file,
and handles multi-page documents. and handles multi-page documents.
@@ -173,10 +173,9 @@ def generate_field_level_conversations(labels_path, image_root, system_prompt_pa
print(f"Formatted data saved to: {output_path}") print(f"Formatted data saved to: {output_path}")
# --- Conversations Generation for only Images --- # --- Conversations Generation for only Images ---
def generate_image_only_conversations(image_root, system_prompt_path, questions_path, output_path): def generate_vq_question(image_root, system_prompt_path, questions_path, output_path):
""" """
Generates conversational VQA pairs for each document based on images only (no labels). Generates conversational VQA pairs for each document based on images only (no labels).
Groups all images with the same prefix (including _1_scale, _2_scale, etc.) into the same conversation.
Each conversation contains a system and user message for each question in the question bank. Each conversation contains a system and user message for each question in the question bank.
""" """
system_prompt = read_text_file(system_prompt_path) system_prompt = read_text_file(system_prompt_path)
@@ -234,4 +233,4 @@ if __name__ == "__main__":
# Run the main generation function # Run the main generation function
# generate_field_level_conversations(LABELS_FILE, IMAGE_ROOT, SYSTEM_PROMPT_FILE, QUESTION_BANK_FILE, ANSWER_BANK_FILE, OUTPUT_FILE) # generate_field_level_conversations(LABELS_FILE, IMAGE_ROOT, SYSTEM_PROMPT_FILE, QUESTION_BANK_FILE, ANSWER_BANK_FILE, OUTPUT_FILE)
generate_image_only_conversations(IMAGE_ROOT, SYSTEM_PROMPT_FILE, QUESTION_BANK_FILE, OUTPUT_FILE) generate_vq_question(IMAGE_ROOT, SYSTEM_PROMPT_FILE, QUESTION_BANK_FILE, OUTPUT_FILE)