Implementation:OpenGVLab InternVL ScienceQA Prompt Conversion

Knowledge Sources	OpenGVLab_InternVL
Domains	Data Preparation, Benchmark, ScienceQA
Last Updated	2026-02-07 14:00 GMT

Overview

This module provides prompt template construction functions for converting ScienceQA benchmark problems into LLaVA-compatible conversation formats.

Description

The convert_sqa_to_llava_base_prompt.py file is a core data preprocessing library used for ScienceQA dataset preparation. It provides helper functions to extract structured fields from ScienceQA problem dictionaries and assemble them into various prompt formats.

Field extraction functions:

• get_question_text: Extracts the question string
• get_context_text: Combines hint text and optional image caption
• get_choice_text: Formats multiple choice options as "(A) ..., (B) ..."
• get_answer: Maps answer index to letter option
• get_lecture_text / get_solution_text: Extracts lecture and solution with newline escaping for GPT-3 compatibility

Prompt assembly functions: Three variants (create_one_example, create_one_example_chatbot, create_one_example_gpt4) build prompts with configurable input-output format strings:

• Input formats: CQM (Context-Question-Multiple choice), QCM, QCML, QCME, QCMLE, and more
• Output formats: A (answer only), AL, AE, ALE, LA, EA, LEA, ELA, LEPA (with lecture, explanation, answer combinations)

Prompt builder functions: build_prompt creates n-shot text prompts, build_prompt_chatbot creates chatbot-style input/output pairs, and build_prompt_gpt4 creates OpenAI-compatible message arrays with system/user/assistant roles.

Usage

Use this module when preparing ScienceQA data for training or evaluating LLaVA models. It is imported by the main conversion script (convert_sqa_to_llava.py) that processes the full dataset.

Code Reference

Source Location

• Repository: OpenGVLab_InternVL
• File: internvl_chat_llava/scripts/convert_sqa_to_llava_base_prompt.py
• Lines: 1-334

Signature

def get_question_text(problem) -> str
def get_context_text(problem, use_caption) -> str
def get_choice_text(problem, options) -> str
def get_answer(problem, options) -> str
def get_lecture_text(problem) -> str
def get_solution_text(problem) -> str

def create_one_example_chatbot(format, question, context, choice, answer, lecture, solution, test_example=True) -> tuple
def create_one_example(format, question, context, choice, answer, lecture, solution, test_example=True) -> str
def create_one_example_gpt4(format, question, context, choice, answer, lecture, solution, test_example=True) -> tuple

def build_prompt_chatbot(problems, shot_qids, prompt_format, use_caption=False, options=["A","B","C","D","E"], is_test=False) -> dict
def build_prompt(problems, shot_qids, test_qid, args) -> str
def build_prompt_gpt4(problems, shot_qids, test_qid, args) -> list

Import

from convert_sqa_to_llava_base_prompt import (
    build_prompt, build_prompt_chatbot, build_prompt_gpt4,
    create_one_example, get_question_text, get_context_text
)

I/O Contract

Inputs

Name	Type	Required	Description
problems	dict	Yes	Dictionary mapping question IDs to ScienceQA problem dicts with keys: question, hint, caption, choices, answer, lecture, solution
shot_qids	list	Yes	List of question IDs to use as few-shot examples
prompt_format	str	Yes	Format string like "QCM-A" or "QCM-LEPA" specifying input-output structure
use_caption	bool	No	Whether to include image captions in context (default: False)
options	list	No	Answer option letters (default: ["A","B","C","D","E"])

Outputs

Name	Type	Description
prompt	str	Formatted prompt string (from build_prompt)
examples	dict	Dictionary of (input, output) pairs keyed by question ID (from build_prompt_chatbot)
prompt_array	list	OpenAI-format message list with role/content dicts (from build_prompt_gpt4)

Usage Examples

Basic Usage

from convert_sqa_to_llava_base_prompt import build_prompt_chatbot

# Build chatbot-style prompts for ScienceQA
problems = json.load(open("problems.json"))
examples = build_prompt_chatbot(
    problems,
    shot_qids=["1", "2", "3"],
    prompt_format="QCM-LEPA",
    use_caption=True
)
# examples["1"] -> (input_str, output_str)

Related Pages

• Principle:OpenGVLab_InternVL_Multimodal_Data_Preparation