generate_response.py

# -*- coding: utf-8 -*-
import math
import time

import nltk
import openai
import re
import os
import json
import subprocess
import argparse
import random
import string
from nltk.corpus import stopwords

# Standard Library Modules
import argparse

# External Modules
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, set_seed

# set random seed
set_seed(42)

from AgentFramework.programmer import programmer_main
from AgentFramework.designer import designer_main
# working on the assumption that executor_main reads from generated files
from AgentFramework.executor import executor_main

B_INST_CLLAMA, E_INST_CLLAMA = "[INST]", "[/INST]"
B_SYS_CLLAMA, E_SYS_CLLAMA = "<<SYS>>\n", "\n<</SYS>>\n\n"
openai.api_key = os.environ['OPENAI_KEY']
PROMPT_START_0 = 'Generate Python3 code (Markdown):\n'
PROMPT_START_1 = 'Generate either Python3 code only (Markdown) or no code:\n'
PROMPT_START_2 = 'Generate either Python3 code only (Markdown) or ask questions:\n'
PROMPT_START_3 = 'You are an expert software developer. Generate Python3 code (code must has Markdown in response) in below information. Alternatively, you can ask clarifying questions: \n'
PROMPT_START_3_v2 = 'You are an expert software developer who writes high quality code. With below information, please either generate Python3 code (Respond directly with code only with markdown), or ask clarifying questions: \n'
# TODO: try this new prompt
PROMPT_START_3_v3 = 'You are an expert software developer who writes high quality code. With below information, please either generate Python3 code (only one code block with markdown in response), or ask clarifying questions (no markdown in response): \n'
PROMPT_START_3_v4 = '\n Based on the information below, you can choose to either generate Python3 code (Respond directly with code only with markdown), or ask clarifying questions. \n'
ORIGINAL_PROMPT_START_0 = 'You are an expert software developer who writes high quality code. With below information, please generate Python3 code (Respond directly with code only with markdown): \n'

PROMPT_EVALUATE_QUESTIONS_V1 = 'The original description of a coding problem is modified so that the requirements become inconsistent, incomplete, or ambiguous. Given the modified description, some clarifying questions were raised to clarify the description. Given the original and modified problem description, evaluate the quality of the questions. Please provide an integer representing the quality of questions (3: Good questions that recover all missing info. 2: Fair questions that recover some missing info. 1: Bad questions or irrelevant content).\n  QUALITY=[your int] \n Please also provide answers to the questions to recover the missing requirements! Be sure to add what is new or different in the original descrpition in your answer, compared with the modified problem description! \n ANSWERS=```[your answer]```  \n Please strictly follow the format QUALITY=[the int] and ANSWERS=```[the answer]``` in the response! Surround your answer with markup! \n\n ### Questions: {clarifying_questions} \n ### Problem Description: {problem} \n ### Original Description: {missing_information} \n'
PROMPT_EVALUATE_QUESTIONS_V2 = 'The original description of a coding problem is modified so that the requirements become inconsistent, incomplete, or ambiguous. Given the modified description, some clarifying questions were raised to clarify the description. Given the original and modified problem description, evaluate the quality of the questions. Please provide an integer representing the quality of questions (3: Good questions that recover all missing info. 2: Fair questions that recover some missing info. 1: Bad questions or irrelevant content).\n  QUALITY=[your int] \n Please also provide answers to the questions to recover the missing requirements! Be sure to add what is new or different in the original descrpition in your answer, compared with the modified problem description! \n ANSWERS=```[your answer]```  \n Please strictly follow the format QUALITY=[the int] and ANSWERS=```[the answer]``` in the response! Surround your answer with markup! \n\n ### Questions: {clarifying_questions} \n ### Problem Description: {problem} \n ### Original Description: {missing_information} \n'
# pretty bad prompt as it returns code in answers...
PROMPT_EVALUATE_QUESTIONS_V3 = 'The original description of a coding problem is modified so that the requirements become incomplete, inconsistent, or ambiguous. Given the modified description, some clarifying questions may be raised to clarify the description. Provide answers to the questions to recover the requirements in the original problem description compared to the modified one. Be sure to return empty answers if there is no valid clarifying question or code with markup! \n ANSWERS=```[your answer]```  \n Please also provide an integer representing the quality of clarifying questions (3: Good questions that recover the modified requirements. 2: Fair questions but they cannot help recover the modified requirements. 1: No valid questions).\n  QUALITY=[your int] \n Please strictly follow the format ANSWERS=```[the answer]``` and QUALITY=[the int] in the response! Surround your answer with markup! \n ### ORIGINAL PROBLEM DESCRIPTION:\n {missing_information} \n ### MODIFIED PROBLEM DESCRIPTION:\n {problem} \n ### CLARIFYING QUESTIONS:\n{clarifying_questions} \n'

# TODO: try this new prompt
# tried once but got no 0 Quality for 5 problems (0-5), so disable this
PROMPT_EVALUATE_QUESTIONS_V4 = 'The original description of a coding problem is modified so that the requirements become inconsistent, incomplete, or ambiguous. Given the modified description, some clarifying questions were raised to clarify the description. Given the original and modified problem description, evaluate the quality of the clarifying questions. Please provide an integer representing the quality of questions (3: Good questions that recover the modified requirements; 2: Fair questions but they cannot help recover the modified requirements; 1: No questions).\n  QUALITY=[your int] \n Please also provide answers to the clarifying questions to recover the modified requirements in the original problem description compared to the modified one. If there is no clarifying questions at all, return empty answers. \n ANSWERS=```[your answer]```  \n Please strictly follow the format QUALITY=[the int] and ANSWERS=```[the answer]``` in the response! Surround your answer with markdown! \n\n ### Questions: Certainly! Can you please provide more specific details on what the `candidate` function needs to do with the input list of strings and the string `x`? This will help me generate the Python3 code according to your requirements. \n ### Modified Problem Description: def candidate(strings: List[str], x: str) -> List[str]:\n    """ Process an input list of strings\n    """ \n ### Original Description: def filter_by_substring(strings: List[str], substring: str) -> List[str]:\n    """ Filter an input list of strings only for ones that contain given substring\n    >>> filter_by_substring([], \'a\')\n    []\n    >>> filter_by_substring([\'abc\', \'bacd\', \'cde\', \'array\'], \'a\')\n    [\'abc\', \'bacd\', \'array\']\n    """ \n ### Response: QUALITY=3\nANSWERS=```The `candidate` function needs to filter the input list of strings such that only strings containing the substring `x` are included in the output list.```\n\n ### Questions: Sure, I can help you with that! Here\'s the Python 3 code for the `candidate` function:\n```python\ndef candidate(num):\n    if num < 0:\n        return (1, 1)\n    else:\n        return (1, num)\n```\n ### Modified Problem Description: def candidate(num):\n    """Example:\n        candidate(-12) ==> (1, 1)\n        candidate(123) ==> (1, 2)\n    """\n ### Original Description: def even_odd_count(num):\n    """Given an integer. return a tuple that has the number of even and odd digits respectively.\n     Example:\n        even_odd_count(-12) ==> (1, 1)\n        even_odd_count(123) ==> (1, 2)\n    """\n ### Response: QUALITY=1\nANSWERS=```no questions```\n\n ### Questions: {clarifying_questions} \n ### Modified Problem Description: {problem} \n ### Original Description: {missing_information} \n ### Response: \n'

PROMPT_EVALUATE_QUESTIONS = 'The original description of a coding problem is modified so that the requirements become inconsistent, incomplete, or ambiguous. Given the modified description, some clarifying questions were raised to clarify the description. Given the original and modified problem description, evaluate the quality of the clarifying questions. Please provide an integer representing the quality of questions (3: Good questions that recover the modified requirements; 2: Fair questions but they cannot help recover the modified requirements; 1: No questions).\n  QUALITY=[your int] \n Please also provide answers to the clarifying questions to recover the modified requirements in the original problem description compared to the modified one. If there is no clarifying questions at all, return empty answers. \n ANSWERS=```[your answer]```  \n Please strictly follow the format QUALITY=[the int] and ANSWERS=```[the answer]``` in the response! Surround your answer with markdown! \n\n ### Questions: {clarifying_questions} \n ### Modified Problem Description: {problem} \n ### Original Description: {missing_information} \n'

PROMPT_2ND_ROUND = '\n Given above conversations, generate Python code directly (Markdown) to solve the coding problem:\n'
OK_PROMPT_CODEGEN = 'Generate Python code directly (Markdown) to solve the coding problem. \n\n'
OK_PROMPT_CLARIFY_Q = 'Given the programming problem, ask clarifying questions if the requirements in the given problem description are incomplete, inconsistent or ambiguous for solving the problem correctly and passing the tests. \n If no need to ask clarifying questions, return strictly \'NO_QUESTIONS\' only. Otherwise, return the clarifying questions. \n\n ### Problem: \n {problem}'
OK_PROMPT_CLARIFY_Q_V1 = 'Given the coding problem description and the generated code above, decide whether to ask clarifying questions that are necessary to solve the problem correctly. \n If no need to ask clarifying questions, return strictly \'NO_QUESTIONS\' only. Otherwise, return the clarifying questions. \n\n'
OK_MODEL = 'gpt-3.5-turbo-0125'

# Instruction-tuned Models and Foundation Models have different nl_2_pl/pl_2_nl prompts and functions
INSTRUCTION_MODELS = [
    "codellama/CodeLlama-7b-Instruct-hf",
    "codellama/CodeLlama-13b-Instruct-hf",
    "codellama/CodeLlama-34b-Instruct-hf",
    "HuggingFaceH4/starchat-beta",
]
FOUNDATION_MODELS = [
    "bigcode/starcoder",
    "bigcode/starcoderplus",
    "bigcode/starcoderbase",
    "bigcode/starcoderbase-7b",
    "bigcode/starcoderbase-3b",
    "bigcode/starcoderbase-1b",
    "codellama/CodeLlama-7b-hf",
    "codellama/CodeLlama-13b-hf",
    "codellama/CodeLlama-34b-hf",
]


# prompt settings
CODELLAMA_NL_2_PL_HUMANEVAL = [
    {  # Instructions
        "role": "system",
        "content": PROMPT_START_3_v2
        + " Note that if you decide to generate code, please respond directly with code only with markdown! You need to return the complete function! Please only return code surrounded by markdown! Don't write down any thought processes!  \n\n",
    },
    {  # One-Shot Example: user input = function signature + problem description in docstring format
        "role": "user",
        "content": 'from typing import List\n\n\ndef has_close_elements(numbers: List[float], threshold: float) -> bool:\n    '
        + '"""Check if in given list of numbers, are any two numbers closer to each other than\n    given threshold.\n    '
        + '>>> has_close_elements([1.0, 2.0, 3.0], 0.5)\n    False\n    '
        + '>>> has_close_elements([1.0, 2.8, 3.0, 4.0, 5.0, 2.0], 0.3)\n    True\n    """\n',
    },
    {  # One-Shot Example: model output = solution
        "role": "assistant",
        "content": '```python\ndef candidate(numbers: List[float], threshold: float) -> bool:\n    for i in range(len(numbers)):\n        for j in range(i+1, len(numbers)):\n            if abs(numbers[i] - numbers[j]) <= threshold:\n                return True\n    return False\n```',
    },
    {  # One-Shot Example: user input = function signature + problem description in docstring format
        "role": "user",
        "content": 'from typing import List\n\n\ndef candidate(...) -> bool:\n \"\"\" Check given a list of number.\"\"\"\n',
    },
    {  # One-Shot Example: model output = solution
        "role": "assistant",
        "content": 'Could you please provide more information on which criteria to check in this function?',
    },
]

CODELLAMA_NL_2_PL_HUMANEVAL_V2 = [
    {  # Instructions
        "role": "system",
        "content": "You are an expert software developer who writes high quality code. Given a coding problem, please either generate Python code, or ask clarifying questions. "
        + "If you decide to generate code: please strictly follow these: Respond directly with code only with markdown! You need to return the complete function! Please only return code surrounded by markdown. Don't write down any thought processes!  \n\n",
    },
]

NL_2_PL_HUMANEVAL = [
    {  # Instructions
        "role": "system",
        "content": "Solve a coding problem in Python. "
        + "Given the function signature and the problem description in the docstring, "
        + "you only need to continue to complete the function body. "
        + "Please strictly follow the format of the example below! "
        + "Don't write down any thought processes! "
        + "Don't copy the problem description! "
        + "You must use correct indentation! "
        + "Make sure your return statement is always inside the function! "
        + "Make sure your output always starts with an indentation of exactly 4 spaces! "
        + "Output an indentation of 4 spaces first before you write anything else! You’d better be sure. \n\n",
    },
    {  # One-Shot Example: user input = function signature + problem description in docstring format
        "role": "user",
        "content": 'from typing import List\n\n\ndef has_close_elements(numbers: List[float], threshold: float) -> bool:\n    '
        + '"""Check if in given list of numbers, are any two numbers closer to each other than\n    given threshold.\n    '
        + '>>> has_close_elements([1.0, 2.0, 3.0], 0.5)\n    False\n    '
        + '>>> has_close_elements([1.0, 2.8, 3.0, 4.0, 5.0, 2.0], 0.3)\n    True\n    """\n',
    },
    {  # One-Shot Example: model output = solution
        "role": "assistant",
        "content": '    sorted_numbers = sorted(numbers)\n    for i in range(len(sorted_numbers) - 1):\n        '
        + 'if sorted_numbers[i + 1] - sorted_numbers[i] < threshold:\n            return True\n    return False\n\n',
    },
    {  # Instructions to emphasize the format
        "role": "system",
        "content": "\nPlease strictly follow the format of the example above! "
        + "You must use correct indentation! "
        + "Make sure your return statement is always inside the function! "
        + "Make sure your output always starts with an indentation of exactly 4 spaces! "
        + "Output an indentation of 4 spaces first before you write anything else! "
        + "You’d better be sure. \n\n",
    },
]

PL_2_NL_HUMANEVAL = [
    {  # Instructions
        "role": "system",
        "content": "Given a Python solution to a coding problem, "
        + "write an accurate problem description for it in the format of Python docstring without 'Args' and 'Returns'. "
        + "Please strictly follow the format of the example below!"
        + "Provide all necessary details to accurately describe the problem, but in a concise way! "
        + "Make sure to give a few examples of inputs and outputs in the docstring! "
        + "Make sure the docstring has no 'Args' and no 'Returns'! "
        + "You can only write a text desciption with a few examples as shown in the example below!  "
        + "Make sure your output always starts with an indentation of exactly 4 spaces! "
        + "You’d better be sure. \n\n",
    },
    {  # One-Shot Example: user input = function signature + candidate solution
        "role": "user",
        "content": 'from typing import List\n\n\ndef has_close_elements(numbers: List[float], threshold: float) -> bool:\n    '
        + 'sorted_numbers = sorted(numbers)\n    for i in range(len(sorted_numbers) - 1):\n        '
        + 'if sorted_numbers[i + 1] - sorted_numbers[i] < threshold:\n            return True\n    return False\n\n',
    },
    {  # One-Shot Example: model output = problem description in docstring format
        "role": "assistant",
        "content": '    """Check if in given list of numbers, are any two numbers closer to each other than\n    '
        + 'given threshold.\n    >>> has_close_elements([1.0, 2.0, 3.0], 0.5)\n    False\n    '
        + '>>> has_close_elements([1.0, 2.8, 3.0, 4.0, 5.0, 2.0], 0.3)\n    True\n    """\n',
    },
    {  # Instructions to emphasize the format
        "role": "system",
        "content": "\nPlease strictly follow the format of the example above! "
        + "Provide all necessary details to accurately describe the problem, but in a concise way! "
        + "Make sure to give a few examples of inputs and outputs in the docstring! "
        + "Make sure the docstring has no 'Args' and no 'Returns'! "
        + "You can only write a text desciption with a few examples as shown in the example above!  "
        + "Make sure your output always starts with an indentation of exactly 4 spaces! "
        + "You’d better be sure. \n\n",
    },
]

NL_2_PL_MBPP = [
    {  # Instructions
        "role": "system",
        "content": "Solve a coding problem in Python. "
        + "Given the function signature and the problem description in the docstring, you only need to continue to complete the function body. "
        + "Please strictly follow the format of the example below! "
        + "Don't write down any thought processes! "
        + "Don't copy the problem description! "
        + "You must use correct indentation! "
        + "Make sure your return statement is always inside the function! "
        + "Make sure your output always starts with an indentation of exactly 4 spaces! "
        + "Output an indentation of 4 spaces first before you write anything else! "
        + "You’d better be sure. \n\n",
    },
    {  # One-Shot Example: user input = function signature + problem description in docstring format
        "role": "user",
        "content": 'def similar_elements(test_tup1, test_tup2):\n    '
        + '""" Write a function to find the shared elements from the given two lists.\n    """\n',
    },
    {  # One-Shot Example: model output = solution
        "role": "assistant",
        "content": '    res = tuple(set(test_tup1) & set(test_tup2))\n    return (res)\n\n',
    },
    {  # Instructions to emphasize the format
        "role": "system",
        "content": "\nPlease strictly follow the format of the example above! "
        + "You must use correct indentation! "
        + "Make sure your return statement is always inside the function! "
        + "Make sure your output always starts with an indentation of exactly 4 spaces! "
        + "Output an indentation of 4 spaces first before you write anything else! "
        + "You’d better be sure. \n\n",
    },
]

PL_2_NL_MBPP = [
    {  # Instructions
        "role": "system",
        "content": "Given a Python solution to a coding problem, write an accurate problem description for it in the format of Python docstring"
        + "Please strictly follow the format of the example below!"
        + "Provide all necessary details to accurately describe the problem, but in a concise way! "
        + "Make sure the docstring has no 'Args', no 'Returns', and no 'Examples'! "
        + "You can only write a plain text desciption as shown in the example below! "
        + "Make sure your output always starts with an indentation of exactly 4 spaces! "
        + "You’d better be sure. \n\n",
    },
    {  # One-Shot Example: user input = function signature + candidate solution
        "role": "user",
        "content": 'def similar_elements(test_tup1, test_tup2):\n    res = tuple(set(test_tup1) & set(test_tup2))\n    return (res)\n\n',
    },
    {  # One-Shot Example: model output = problem description in docstring format
        "role": "assistant",
        "content": '    """ Write a function to find the shared elements from the given two lists.\n    """\n',
    },
    {  # Instructions to emphasize the format
        "role": "system",
        "content": "\nPlease strictly follow the format of the example above! "
        + "Provide all necessary details to accurately describe the problem, but in a concise way! "
        + "Make sure the docstring has no 'Args', no 'Returns', and no 'Examples'! "
        + "You can only write a plain text desciption as shown in the example above! "
        + "Make sure your output always starts with an indentation of exactly 4 spaces! "
        + "You’d better be sure. \n\n",
    },
]

ONE_SHOT_HUMANEVAL = (
    'def has_close_elements(numbers: List[float], threshold: float) -> bool:\n    '
    + '"""Check if in given list of numbers, are any two numbers closer to each other than\n    given threshold.\n    '
    + '>>> has_close_elements([1.0, 2.0, 3.0], 0.5)\n    False\n    >>> has_close_elements([1.0, 2.8, 3.0, 4.0, 5.0, 2.0], 0.3)\n    '
    + 'True\n    """\n    sorted_numbers = sorted(numbers)\n    for i in range(len(sorted_numbers) - 1):\n        '
    + 'if sorted_numbers[i + 1] - sorted_numbers[i] < threshold:\n            return True\n    return False\n\n'
)

ONE_SHOT_MBPP = (
    'def similar_elements(test_tup1, test_tup2):\n    """ Write a function to find the shared elements from the given two lists.\n    '
    + '"""\n    res = tuple(set(test_tup1) & set(test_tup2))\n    return (res)\n\n'
)


def generate_text(model, tokenizer, prompt_text, args, eos_token_id=None):
    if eos_token_id is None:
        eos_token_id = tokenizer.eos_token_id
    model_inputs = tokenizer(prompt_text, padding=False, add_special_tokens=False, return_tensors="pt")
    model_inputs["prompt_text"] = prompt_text
    input_ids = model_inputs["input_ids"]
    attention_mask = model_inputs.get("attention_mask", None)
    # Allow empty prompts
    if input_ids.shape[1] == 0:
        input_ids = None
        attention_mask = None
        in_b = 1
    else:
        in_b = input_ids.shape[0]
    # BS x SL
    if args.gen_length is None:
        max_length = args.seq_length
    else:
        max_length = input_ids.shape[1] + args.gen_length
    print("Generating text...")
    generated_sequence = model.generate(
        input_ids=input_ids.to(model.device),
        attention_mask=attention_mask.to(model.device),
        temperature=args.temperature,
        top_k=args.top_k,
        top_p=args.top_p,
        max_length=max_length,
        do_sample=args.do_sample,
        num_beams=args.num_beams,
        num_return_sequences=args.num_return_sequences,
        pad_token_id=tokenizer.pad_token_id,
        eos_token_id=eos_token_id,
    )
    out_b = generated_sequence.shape[0]
    generated_sequence = generated_sequence.reshape(in_b, out_b // in_b, *generated_sequence.shape[1:])
    generated_sequence = generated_sequence[0].cpu().numpy().tolist()
    records = []
    print("Decoding text...")
    for sequence in generated_sequence:
        text = tokenizer.decode(sequence, skip_special_tokens=True)
        prompt_length = len(tokenizer.decode(input_ids[0], skip_special_tokens=True))
        all_text = text[prompt_length:]
        record = {"generated_text": all_text}
        records.append(record)
    return records


def get_completion_starchat_nl_to_pl(prompt, user_input, model, tokenizer, args):
    # select the correct in-context learning prompt based on the task
    messages = prompt + [{"role": "user", "content": user_input}]
    try:
        dialogue_template = DialogueTemplate.from_pretrained(args.model_name_or_path)
    except Exception:
        print("No dialogue template found in model repo. Defaulting to the `no_system` template.")
        dialogue_template = get_dialogue_template("no_system")
    dialogue_template.messages = messages
    formatted_prompt = dialogue_template.get_inference_prompt_nl_to_pl()
    # Debug
    # print(formatted_prompt)
    # get completion from code lm
    output = generate_text(
        model,
        tokenizer,
        formatted_prompt,
        args,
        eos_token_id=tokenizer.convert_tokens_to_ids(dialogue_template.end_token),
    )
    completion = output[0]["generated_text"]
    # post-processing
    completion_lines = completion.split("\n")
    processed_completion = ""
    for line in completion_lines:
        if line.startswith("    "):
            processed_completion += line + "\n"
    # remove extra docstring
    # find all occurrences of three consecutive double quotes
    res = [i for i in range(len(processed_completion)) if processed_completion.startswith('"""', i)]
    # if res is empty, check for both single quotes
    if not res:
        res = [i for i in range(len(processed_completion)) if processed_completion.startswith("'''", i)]
    # if found an extra docstring, remove it
    if res:
        # get end position of the extra docstring, remove everything before it
        try:
            end_position = res[1] + 3
            processed_completion = processed_completion[end_position:]
            if processed_completion.startswith("\n"):
                processed_completion = processed_completion[1:]
        except IndexError:
            pass

    # Debug
    print(processed_completion)
    return processed_completion


def get_completion_starchat_pl_to_nl(prompt, user_input, model, tokenizer, args):
    # select the correct in-context learning prompt based on the task
    messages = prompt + [{"role": "user", "content": user_input}]
    try:
        dialogue_template = DialogueTemplate.from_pretrained(args.model_name_or_path)
    except Exception:
        print("No dialogue template found in model repo. Defaulting to the `no_system` template.")
        dialogue_template = get_dialogue_template("no_system")
    dialogue_template.messages = messages
    formatted_prompt = dialogue_template.get_inference_prompt_pl_to_nl()
    # Debug
    # print(formatted_prompt)
    # get completion from code lm
    output = generate_text(
        model,
        tokenizer,
        formatted_prompt,
        args,
        eos_token_id=tokenizer.convert_tokens_to_ids(dialogue_template.end_token),
    )
    completion = output[0]["generated_text"]
    # post-processing: extract the docstring
    completion = completion.replace("python", "")
    completion_parts = completion.split("```")
    if len(completion_parts) > 1:
        completion = completion_parts[1]
    completion_parts = completion.split('"""')
    if len(completion_parts) > 1:
        completion = completion_parts[1]
    completion_parts = completion.split("'''")
    if len(completion_parts) > 1:
        completion = completion_parts[1]

    # double check
    if not completion.startswith('"""'):
        completion = '"""' + completion
    if not completion.endswith('"""'):
        completion = completion + '\n"""'
    completion_lines = completion.split("\n")
    for idx, line in enumerate(completion_lines):
        if not line.startswith("    "):
            completion_lines[idx] = "    " + line
    completion = "\n".join(completion_lines)
    # Debug
    print(completion)
    return completion


def get_completion_codellama_instruct_nl_to_pl(
    prompt, user_input, model, tokenizer, args
):  # reference: https://github.com/facebookresearch/codellama/blob/main/llama/generation.py
    
    formatted_prompt = ""
    # the template is being used
    if prompt == '':
        formatted_prompt = user_input 
    else:
    # select the correct in-context learning prompt based on the task
        messages = [{"role": "user", "content": user_input}]
    
        for msg in messages:
            if msg["role"] == "user":
                content = msg["content"].strip()
                formatted_prompt += tokenizer.bos_token + f"{B_INST_CLLAMA} " + content + f" {E_INST_CLLAMA} "
            elif msg["role"] == "assistant":
                formatted_prompt += " " + msg["content"].strip() + " " + tokenizer.eos_token
            # system prompt doesn't work well for Code Llama-Instructs
            elif msg["role"] == "system":
                formatted_prompt += f"{B_SYS_CLLAMA}" + msg["content"] + f"{E_SYS_CLLAMA}"

    # Debug
    print('\nformatted_prompt:\n',formatted_prompt)
    
    output = generate_text(model, tokenizer, formatted_prompt, args)
    completion = output[0]["generated_text"]

    print('\ncompletion:\n',completion)
    return completion

    # post-processing
    #completion_lines = completion.split("\n")
    #processed_completion = ""
    #for line in completion_lines:
    #    if line.startswith("    "):
    #        processed_completion += line + "\n"

    # remove extra docstring
    # find all occurrences of three consecutive double quotes
    #res = [i for i in range(len(processed_completion)) if processed_completion.startswith('"""', i)]
    # if res is empty, check for both single quotes
    #if not res:
    #    res = [i for i in range(len(processed_completion)) if processed_completion.startswith("'''", i)]
    # if found an extra docstring, remove it
    #if res:
        # get end position of the extra docstring, remove everything before it
    #    try:
    #        end_position = res[1] + 3
    #        processed_completion = processed_completion[end_position:]
    #        if processed_completion.startswith("\n"):
    #            processed_completion = processed_completion[1:]
    #    except IndexError:
    #        pass

    # Debug
    #print(processed_completion)
    #return processed_completion


def get_completion_codellama_instruct_pl_to_nl(prompt, user_input, model, tokenizer, args):
    # select the correct in-context learning prompt based on the task
    messages = prompt + [{"role": "user", "content": user_input}]
    formatted_prompt = ""
    for msg in messages:
        if msg["role"] == "user":
            if args.input_path.endswith("EvalPlus-Mini-v0.1.6_reformatted.jsonl"):
                content = (
                    msg["content"]
                    + "\n\nWhat should be the docstring of the above function? Please only write down the docstring with some examples."
                )
            elif args.input_path.endswith("MBPP-S_test_reformatted.jsonl"):
                content = (
                    msg["content"]
                    + "\n\nWhat should be the docstring of the above function? Please write down the docstring only in words without any examples!"
                )
            else:
                raise ValueError(f"Input file {args.input_path} not supported")
            formatted_prompt += tokenizer.bos_token + "[INST] " + content + " [/INST] "
        elif msg["role"] == "assistant":
            formatted_prompt += " " + msg["content"].strip() + " " + tokenizer.eos_token
    # Debug
    # print(formatted_prompt)
    output = generate_text(model, tokenizer, formatted_prompt, args)
    completion = output[0]["generated_text"]
    # post-processing
    completion_lines = completion.split("\n")
    for idx, line in enumerate(completion_lines):
        if not line.startswith("    "):
            completion_lines[idx] = "    " + line.lstrip()
    completion = "\n".join(completion_lines)
    # add docstring guards if not present
    if completion.startswith('    """') and not completion.endswith('"""'):
        completion = completion + '"""'
    elif completion.startswith("    '''") and not completion.endswith("'''"):
        completion = completion + "'''"
    # Debug
    print(completion)
    return completion


def get_completion_codellama(
    prompt, user_input, model, tokenizer, args
):  # prompt is ONE_SHOT_HUMANEVAL or ONE_SHOT_MBPP
    user_input = prompt + user_input
    output = generate_text(model, tokenizer, user_input, args)
    completion = output[0]["generated_text"]
    # post-processing: extract the function body
    processed_completion = ""
    completion_lines = completion.split("\n")
    for line in completion_lines:
        if line.startswith("    "):
            processed_completion += line + "\n"
        else:
            break
    # Debug
    print(processed_completion)
    return processed_completion


def get_completion_codellama_fim(
    prompt, function_signature, function_body, model, tokenizer, args
):  # prompt is ONE_SHOT_HUMANEVAL or ONE_SHOT_MBPP
    function_signature = prompt + function_signature
    fim_prompt = ""
    # check indent
    fim_prompt += (
        " <PRE>" + function_signature + "    \"\"\"\n    " + " <SUF>" + "    \"\"\"\n" + function_body + " <MID>"
    )
    # Debug
    # print(fim_prompt)
    output = generate_text(model, tokenizer, fim_prompt, args)
    completion = output[0]["generated_text"]
    completion = "    \"\"\"\n    " + completion + "    \"\"\"\n"
    # Debug
    print(completion)
    return completion


def get_completion_starcoder(
    prompt, user_input, model, tokenizer, args
):  # prompt is ONE_SHOT_HUMANEVAL or ONE_SHOT_MBPP
    user_input = prompt + user_input
    output = generate_text(model, tokenizer, user_input.strip(), args)
    completion = output[0]["generated_text"]
    # post-processing: extract the function body
    processed_completion = ""
    completion_lines = completion.split("\n")
    start = False
    for line in completion_lines:
        if line.startswith("    "):
            start = True
            processed_completion += line + "\n"
        else:
            if start:
                break
    # Debug
    print(processed_completion)
    return processed_completion


def get_completion_starcoder_fim(
    prompt, function_signature, function_body, model, tokenizer, args
):  # prompt is ONE_SHOT_HUMANEVAL or ONE_SHOT_MBPP
    function_signature = prompt + function_signature
    fim_prompt = ""
    # TODO check indent
    fim_prompt += (
        "<fim_prefix>"
        + function_signature
        + "    \"\"\""
        + "<fim_suffix>"
        + "\"\"\"\n"
        + function_body
        + "<fim_middle>"
    )
    output = generate_text(model, tokenizer, fim_prompt, args)
    completion = output[0]["generated_text"]
    completion = "    \"\"\"" + completion + "\"\"\"\n"
    # Debug
    print(completion)
    return completion



def evaluate_clarifying_questions(
    missing_information='',
    clarifying_questions='',
    problem=''
):
    print("\n\n!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!", file=print_file)
    print('!!!!!!! 2nd evaluate_clarifying_questions START !!!!!!!!!!!', file=print_file)
    topn = 1
    temperature = 1.0
    model = 'gpt-3.5-turbo-0125' #'gpt-3.5-turbo'
    content = PROMPT_EVALUATE_QUESTIONS.format(
                missing_information=missing_information,
                clarifying_questions=clarifying_questions,
                problem=problem
            )
    completion = openai.ChatCompletion.create(
        model=model,
        n=topn,
        temperature=temperature,
        messages=[{
            "role": "user",
            "content": content,
        }]
    )
    print('!!!!!!!PROMPT_EVALUATE_QUESTIONS='+content, file=print_file)
    print('!!!!!!!Completion='+completion['choices'][0]['message']['content'], file=print_file)
    # Convert completion content to a string if it's not already a string
    completion_content = str(completion['choices'][0]['message']['content'])

    # Use re.findall() with the completion content
    question_quality = re.findall(r'QUALITY\s*=?\s*(\d+)', completion_content)
    answers = re.findall(r'ANSWERS\s*=?\s*```(.+?)```', completion_content, flags=re.DOTALL)
    answer_str = answers[0] if answers else ""
    question_quality_str = question_quality[0] if question_quality else ""
    print('!!!!!!!answer_str',answer_str, file=print_file)
    print('!!!!!!!question_quality_str',question_quality_str, file=print_file)
    
    print('!!!!!!! 2nd evaluate_clarifying_questions END !!!!!!!!!!!', file=print_file)
    print("!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!\n\n", file=print_file)
    return answer_str, question_quality_str

def create_prompt(description, option='original', percentage=0):
    if option == 'original':
        prompt = PROMPT_START_3
        return prompt + description
    elif option.startswith('randRemove'):
        return PROMPT_START_3 + split_and_remove_chunk(description, percentage)
    elif option.startswith('manualRemove'): # TODO(jwu): WIP
        return PROMPT_START_3_v2 + description
    else:
        return PROMPT_START_3 + split_and_replace_with_random_words(description, percentage)

# A larger vocabulary of common English words
common_words = [
    "apple", "banana", "chocolate", "dog", "elephant", "flower", "guitar",
    "happiness", "internet", "jazz", "kangaroo", "laughter", "mountain", "ocean",
    "penguin", "question", "rainbow", "sunshine", "umbrella", "victory", "wonderful",
    "xylophone", "zebra", "car", "house", "book", "computer", "chair", "table",
    "phone", "shoes", "music", "friend", "beach", "movie", "cake", "coffee",
    "game", "travel", "nature", "art", "garden", "party", "smile", "love", "star",
    "moon", "bird", "child", "family", "money", "dream", "time", "water", "fire",
    "food", "work", "school", "world", "health", "peace", "joy", "knowledge", "color",
    "forest", "planet", "song", "heart", "adventure", "freedom", "success", "history"
]

# Function to generate a random word from the larger vocabulary
def generate_random_common_word():
    return random.choice(common_words)

def split_and_replace_with_random_words(text, percentage):
    # Split the input string into words
    words = text.split()

    # Calculate the length of the input text
    num_words = len(words)

    # Calculate the number of words to replace (approximately one-third)
    num_words_to_replace = calculate_percentage_integer(num_words, percentage)

    if num_words_to_replace == 0:
        return text  # If there are too few words to replace, return the original text

    # Choose a random starting index for the chunk to replace
    start_index = random.randint(0, num_words - num_words_to_replace)

    # Determine the end index of the chunk
    end_index = start_index + num_words_to_replace

    # Generate random common words to replace the chunk
    random_common_words = [generate_random_common_word() for _ in range(num_words_to_replace)]

    # Create a list to store the modified words
    modified_words = words[:start_index] + random_common_words + words[end_index:]

    # Join the modified words to create the modified text
    modified_text = ' '.join(modified_words)

    return modified_text

def split_and_remove_chunk(text, percentage):
    # Split the input string into words
    words = text.split()

    # Calculate the length of the input text
    num_words = len(words)

    # Calculate the number of words to remove (approximately one-third)
    num_words_to_remove = calculate_percentage_integer(num_words, percentage)

    if num_words_to_remove == 0:
        return text  # If there are too few words to remove, return the original text

    # Choose a random starting index for the chunk to remove
    start_index = random.randint(0, num_words - num_words_to_remove)

    # Determine the end index of the chunk
    end_index = start_index + num_words_to_remove

    # Create a list to store the words to keep
    words_to_keep = []

    # Iterate through the words and add them to the list if they are outside the chunk
    for index, word in enumerate(words):
        if index < start_index or index >= end_index:
            words_to_keep.append(word)

    # Join the words to create the modified text
    modified_text = ' '.join(words_to_keep)

    return modified_text

def calculate_percentage_integer(value, percentage):
    # Calculate the result as a floating-point number
    result = value * (percentage / 100.0)
    
    # Round the result to the nearest integer
    rounded_result = round(result)
    
    return rounded_result

# legacy code (randRemove) where only one-round evaluation is enabled
def description_2_code_one_round(prompt, model, topn, temperature, args, open_source_model, tokenizer):
    if model=='comm':
        completion = openai.ChatCompletion.create(
            model='gpt-3.5-turbo',
            n=1,
            temperature=temperature,
            messages=[{"role": "user",
                       "content": prompt},
                      ]
        )
        first_response_list = []
        for i in completion['choices']:
            first_response_list.append(i['message']['content'])

        new_prompt = "You are an expert in software engineering. You will be given the problem description and current code of a coding task. You will decide whether to ask clarifying questions or return the code with markup. \n ### Problem Description: \n"+ prompt + "\n ### Generated Code From Previous Iteration:\n" + first_response_list[0]
        
        completion = openai.ChatCompletion.create(
            model='gpt-3.5-turbo',
            n=topn,
            temperature=temperature,
            messages=[{"role": "user",
                       "content": new_prompt},
                      ]
        )
        response_list = []
        # code_list = []
        for i in completion['choices']:
            response_list.append(i['message']['content'])

    else:
        messages=[{"role": "user", "content": prompt}]
        response_list = generate_response(model, messages, topn, temperature, args, open_source_model, tokenizer)
    code_list = []
    qq_list = []
    for i in range(len(response_list)):
        code = response_2_code(response)
        code_list.append(code)
        qq_list.append('0')
    return response_list, code_list, qq_list

def generate_response_str(model, msgs, temperature, args, open_source_model, tokenizer):
    response_list = generate_response(model, msgs, 1, temperature, args, open_source_model, tokenizer)
    return response_list[0]
    
def generate_response(model, msgs, topn, temperature, args, open_source_model, tokenizer, user_input_without_prompt = '', prompt = ''):
    response_list = []
    if args.model.startswith('starcoder'):
        user_input = tokenizer.apply_chat_template(msgs, tokenize=False)
        for i in range(topn):
            response_list.append(get_completion_starcoder('', user_input, open_source_model, tokenizer, args))
        return response_list        
    elif 'Llama' in args.model or 'deepseek' in args.model or 'CodeQwen' in args.model:
        user_input = tokenizer.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True, return_tensors="pt")
        for i in range(topn):
            if 'two-shot' in args.model:
                response_list.append(get_completion_codellama_instruct_nl_to_pl(CODELLAMA_NL_2_PL_HUMANEVAL, user_input_without_prompt, open_source_model, tokenizer, args))
            elif 'Llama' in args.model:
                CODELLAMA_NL_2_PL_PROMPT = [
                    {  # Instructions
                        "role": "system",
                        "content": prompt,
                    },
                ]
                response_list.append(get_completion_codellama_instruct_nl_to_pl('CODELLAMA', prompt + user_input_without_prompt, open_source_model, tokenizer, args))
            else:
                response_list.append(get_completion_codellama_instruct_nl_to_pl('', user_input, open_source_model, tokenizer, args))
        return response_list
    elif model == 'Okanagan':
        # this code assume topn=1
        # set the real model used by Okanagan
        coder_response = generate_response_str(OK_MODEL, msgs, temperature, args, open_source_model, tokenizer)

        # Reflection
        reflect_messages = [{"role": "user","content": OK_PROMPT_CLARIFY_Q.format(code=coder_response, problem=user_input_without_prompt)}]
        # messages.append({"role": "assistant","content": coder_response})
        # messages.append({"role": "user","content": OK_PROMPT_CLARIFY_Q})
        communicator_response = generate_response_str(OK_MODEL, reflect_messages, temperature, args, open_source_model, tokenizer)
        print("!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!", file=print_file)
        print("!!!!!!!!!!!!!!! Okanagan !!!!!! communicator_response: \n" + communicator_response, file=print_file)
        print("!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!\n", file=print_file)
        #messages.append({"role": "assistant","content": communicator_response})
        if  re.search('no_questions', communicator_response, re.IGNORECASE):
            response_list.append(coder_response)
        else:
            response_list.append(communicator_response)    
        return response_list  
    elif model == 'AgentCoder':

        task_id = msgs[0]["task_id"]
        task_id = task_id.replace("/", "_")
        
        print("Running Programmer")
        responses = programmer_main(model, "python", msgs, openai.api_key, task_id)

        if msgs[0]["clarity_prompt"]=="":
            # no clarifying questions being generated through the prompt
            print("Running Designer")
            test_cases = designer_main(model, "python", responses, openai.api_key, task_id)
            
            print("Running Executor")
            results = executor_main(task_id)
            
            response_list.append(str(results[0]['completion']))
        else:
            # we have asked the model to generate clarifying questions
            response_list.append(str(responses[0]['completion_list']))
        return response_list
    else:
        completion = openai.ChatCompletion.create(
            model=model,
            n=topn,
            temperature=temperature,
            messages=msgs
        )
        for i in completion['choices']:
            response_list.append(i['message']['content'])
        return response_list

def description_2_code_multi_rounds(prompt_modified, task_id, entry_point, prompt, user_input, original_prompt, model, topn, temperature, args, open_source_model, tokenizer, cached_response, cached_qq, cached_answer):
    
    messages = []
    response_list = []
    model_2nd_round = OK_MODEL if model == 'Okanagan' else model
    # ROUND 1
    if model == "AgentCoder":
        # Adding the following: entry_point, task_id, original_prompt for AgentCoder
        if prompt_modified == False:
            messages.append({"task_id": task_id,"prompt": original_prompt, "entry_point": entry_point, "clarity_prompt": ""})
        else:
            messages.append({"task_id": task_id,"prompt": original_prompt, "entry_point": entry_point, "clarity_prompt": PROMPT_START_3_v4})
    else:
        ## 1st round: initial code generation
        full_prompt = OK_PROMPT_CODEGEN + user_input if model == 'Okanagan' else prompt + user_input
        
        print("\n\n!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!", file=print_file)
        print('!!!!!!!!!!!!! prompt:\n' + full_prompt, file=print_file)
        print("!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!\n", file=print_file)
        
        messages.append({"role": "user","content": full_prompt})
    if args.log_phase_output >= 2:
        response_list.append(cached_response)
    else:
        response_list = generate_response(model, messages, topn, temperature, args, open_source_model, tokenizer, user_input, prompt)
        
    if args.log_phase_output == 1:
        return response_list, [], [], []

    code_list = []
    qq_list = []
    ans_list = []
    for i in range(len(response_list)):
        response = response_list[i]
        code = response_2_code_if_no_text(response)
        
        print("\n\n!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!", file=print_file)
        print('!!!!!!!!!!!!! 1st CodeLLM response:\n' + response, file=print_file)
        print('!!!!!!!!!!!!! 1st CodeLLM response code:\n' + code, file=print_file)
        print("!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!\n\n", file=print_file)
        question_quality = '0'
        answer = ''
        if code == '':
            ## 2nd round: question & answer round
            # use LLM-based Evaluator to
            # 1) generate answer,
            # 2) evaluate quality of clarifying questions,
            # 3) generate new code with Q&A
            if args.log_phase_output >= 3:
                answer = cached_answer
                question_quality = cached_qq
            else:
                answer, question_quality = evaluate_clarifying_questions(original_prompt,response,user_input)
            
            if args.log_phase_output == 2:
                ans_list.append(answer)
                qq_list.append(question_quality)
                continue

            ## 3rd round: generate final code: generate 2nd-round code with chat history (Q&A)
            if model == "AgentCoder":
                print("This is the original message:", messages)
                # We can only send one prompt to AgentCoder for now. Adding multiple roles requires major code changes in the original AgentCoder repo
                new_prompt = "Original Question: " + original_prompt + " First Response: " + response + " Feedback: " + answer + " " + PROMPT_2ND_ROUND
                messages[-1]["prompt"] = new_prompt
                for message in messages:
                    message['clarity_prompt'] = ""

                msgs_i = messages.copy()

            else:
                msgs_i = messages.copy()
                msgs_i.append({"role":"assistant","content": response})
                msgs_i.append({"role":"user","content": answer + PROMPT_2ND_ROUND})
            
            response_2nd = generate_response(model_2nd_round, msgs_i, 1, temperature, args, open_source_model, tokenizer)
            code = response_2_code(response_2nd[0])
            
            print("\n\n!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!", file=print_file)
            print('!!!!!!!!!!!!! 3rd CodeLLM input messages:\n', msgs_i, file=print_file)
            print('!!!!!!!!!!!!! 3rd CodeLLM response:\n', response_2nd, file=print_file)
            print("!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!\n\n", file=print_file)
        qq_list.append(question_quality)
        code_list.append(code)
        ans_list.append(answer)
    return response_list, code_list, qq_list, ans_list

def get_ith_element(input_string, i):
    # Split the input string by '_' to create a list of elements
    elements = input_string.split('_')

    # Check if i is a valid index within the range of elements
    if 0 <= i < len(elements):
        return elements[i]
    else:
        return ""  # Return "" if i is out of range

def string_to_int(input_string):
    try:
        result = int(input_string)
        return result
    except ValueError:
        return None  # Return None if the string cannot be converted to an integer

# Return the first triple code snippet. 
def response_2_code(response):
    code_template = re.compile('```.*\n([\s\S]+?)\n```', re.M)
    code = code_template.findall(response)
    if len(code) > 0:
        return code[0] # code[-1] is the last triple code snippet
    else:
        return ''

# returns code only if the response consists solely of code with markups
def response_2_code_if_no_text(response):
    # Adjusted regular expression to allow optional surrounding whitespace
    code_template = re.compile(r'^\s*```.*?\n([\s\S]+?)\n```\s*$', re.M)
    match = code_template.match(response) # fullmatch should be used, but this is currently ok
    if match:
        return match.group(1)  # Return the code block (group 1)
    return ''  # Return empty string if no match is found

def HumanEval_experiment(dataset, dataset_loc, option, model, topn, temperature, args, open_source_model, tokenizer):
    remove_percentage = 0
    log_file = ''
    if option == 'original':
        log_file = './log/dataset_%s_model_%s_topn_%s_temperature_%s.log_%s' % \
                   (dataset, model, topn, temperature, str(args.log_phase_input))
    else:
        log_file = './log/%s_dataset_%s_model_%s_topn_%s_temperature_%s.log_%s' % \
                   (option, dataset, model, topn, temperature, str(args.log_phase_input))
        remove_percentage = string_to_int(get_ith_element(option, 1))
    
    # write printed output to a file (print_file)
    print_file_str = './log/print' + log_file[5:]
    global print_file
    print_file = open(print_file_str, 'a') # append new content if exists already
    
    problem_list = []
    line_cnt = 0
    with open(dataset_loc, 'r') as f:
        for line in f.readlines():
            if args.min_problem_idx < 0 or line_cnt >= args.min_problem_idx:
                problem_list.append(json.loads(line))
            # added by JW
            line_cnt += 1
            if args.max_num_problems >= 0 and line_cnt >= args.max_num_problems:
                break
    # names with prompt type (e.g. 'HumanEval/X_promptX')
    cached_names = set()
    cached_responses = {}
    cached_answers = {}
    cached_qqs = {}
    if os.path.exists(log_file):
        with open(log_file, 'r') as f:
            for line in f:
                content = json.loads(line)
                key = content['name']+'_'+content['prompt_type']
                cached_names.add(key)
                cached_responses[key] = content['response']
                cached_answers[key] = content['answer']
                cached_qqs[key] = content['question_quality']

    response_list = []
    for problem in problem_list:
        print('----------------------problem name: %s--------------------------------' % (problem['name']), flush=True)
        print('using %s to generate response' % (model), flush=True)
        
        if dataset == "HumanEvalComm":
            input_prompt_fields = ['prompt1a','prompt1c','prompt1p','prompt2ac','prompt2ap','prompt2cp','prompt3acp']
        else:  
            input_prompt_fields = ['prompt']
        
        for input_prompt in input_prompt_fields:
            if input_prompt not in problem:
                continue
            key = problem['name'] + '_' + input_prompt
            if args.log_phase_input == args.log_phase_output and key in cached_names:
                continue
            print("********************************************************************", file=print_file)
            print("****** new problem (name="+problem['name']+" input_prompt="+input_prompt+") ******", file=print_file)
            print("********************************************************************", file=print_file)
            description = problem[input_prompt]
            try:
                prompt = create_prompt(description, option, remove_percentage)
                if option.startswith('randRemove'):
                    # legacy part
                    # Dont need to make any changes to this I think, since this is legacy part
                    response_list, code_list, qq_list = description_2_code_one_round(prompt, model, topn, temperature, args, open_source_model, tokenizer)
                else:
                    original_prompt = problem['prompt']
                    entry_point = problem['entry_point']
                    task_id = problem['name']
                    # prompt_start = ORIGINAL_PROMPT_START_0 if input_prompt == 'prompt' else PROMPT_START_3_v2
                    # We will use "prompt_modified" to check whether AgentCoder is getting a modified prompt or an original prompt, based on which, we decide whether to send in a "generate clarifying questions" prompt or not.
                    # A new prompt called PROMPT_START_3_v4 has been created for the same.
                    prompt_modified = False if input_prompt == 'prompt' else True
                    prompt_start = ORIGINAL_PROMPT_START_0 if input_prompt == 'prompt' else PROMPT_START_3_v2
                    response_list, code_list, qq_list, ans_list = description_2_code_multi_rounds(prompt_modified, task_id, entry_point, prompt_start, description, original_prompt, model, topn, temperature, args, open_source_model, tokenizer, cached_responses.get(key, ''), cached_qqs.get(key, 0), cached_answers.get(key, ''))
            except Exception as e:
                print('%s---------%s' % (problem['name'], e), flush=True)
                continue
            for i in range(len(response_list)):
                if args.log_phase_output >= 1:
                    res = {
                        'key': key,
                        'name': problem['name'],
                        'prompt_type': input_prompt,
                        'index': i,
                        'response': response_list[i],
                        'answer': ans_list[i] if i < len(ans_list) else '',
                        'question_quality': qq_list[i] if i < len(qq_list) else '0',
                        'code': code_list[i] if i < len(code_list) else '',
                    }
                    print('response %s is writting into file' % (i), flush=True)
                    json_str = json.dumps(res)

                    # Find the last occurrence of '.log_' in the string
                    last_index = log_file.rfind('.log_')
                    # Remove the substring from the last occurrence of '.log_' to the end, then add new suffix
                    log_file_output = log_file[:last_index] + '.log_' + str(args.log_phase_output)
                    
                    with open(log_file_output, 'a') as f:
                        f.write(json_str + '\n')
                else:
                    res = {
                        'name': problem['name'],
                        'index': i,
                        'response': response_list[i],
                        'original_prompt': description,
                        'modified_prompt': prompt,
                        'prompt_type': input_prompt,
                        'code': code_list[i],
                        'question_quality': qq_list[i],
                        'answer': ans_list[i],
                    }
                    print('response %s is writting into file' % (i), flush=True)
                    json_str = json.dumps(res)
                    with open(log_file, 'a') as f:
                        f.write(json_str + '\n')
            print('%s finish!' % (problem['name']), flush=True)
            # stop with 1 prompt for debugging
            #break
    print('Done!', flush=True)

def test_starcoder(tokenizer, model, user_input, max_length):
    device = "cuda" if torch.cuda.is_available() else "cpu"
    inputs = tokenizer.encode(
        input_ids=user_input,
        # compute a + b 
        #"def print_hello_world():", 
        return_tensors="pt",
        ).to(device)
    print('device=', device)
    outputs = model.generate(
        inputs,
        max_length=max_length,
        )
    print('!!!!!!!!!!')
    print(tokenizer.decode(outputs[0]))
    print('!!!!!!!!!!')

def test_codellama(tokenizer, model, user_input, max_length):
    timea = time.time()
    input_ids = tokenizer(user_input, return_tensors="pt")["input_ids"].to(model.device)
    generated_ids = model.generate(input_ids, max_new_tokens=max_length)
    filling = tokenizer.batch_decode(generated_ids[:, input_ids.shape[1]:], skip_special_tokens = True)[0]
    
    print('!!!!!!!!!!')
    print(filling)
    print('!!!!!!!!!!')
    print("timea = time.time()",-timea + time.time())


# call LLM to generate results from problems
# input: HumanEval.jsonl
# output: file in folder log/
if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument(
        "-d",
        "--dataset",
        type=str,
        choices=['APPS', 'code_contest', 'HumanEval', 'HumanEvalComm'],
        help="Choose dataset",
        required=True,
    )
    parser.add_argument(
        "-m",
        "--model",
        help="Openai Model",
        type=str,
        required=True,
    )
    parser.add_argument(
        "-n",
        "--topn",
        type=int,
        help="Top N candidates",
        required=True,
        default=1,
    )
    parser.add_argument(
        "-t",
        "--temperature",
        type=float,
        help="Set the temperature",
        required=True,
        default='1',
    )
    parser.add_argument(
        "-o",
        "--option",
        type=str,
        #choices=['original', 'rephrase', 'extractive_summarize', 'abstractive_summarize'],
        help="Choose the mode of the experiment",
        required=True,
        default='original'
    )
    parser.add_argument(
        "-s", # legacy
        "--log_phase_input",
        choices=[0,1,2,3],
        type=int,
        help="If not 0, this split the process into phase 1 (1st round LLM response),2 (2nd, answers to questions),3 (3rd, final code generation given chat history). This is name of input log file",
        default=0
    )
    parser.add_argument(
        "-so",
        "--log_phase_output",
        choices=[0,1,2,3],
        type=int,
        help="If not 0, this split the process into phase 1 (1st round LLM response),2 (2nd, answers to questions),3 (3rd, final code generation given chat history). This is name of output log file",
        default=0
    )
    parser.add_argument(
        "-maxp",
        "--max_num_problems",
        type=int,
        help="Max number of problems to run", # limit the number of problems to be run and call ChatGPT. -1 means no such limit
        default=-1,
    )
    parser.add_argument(
        "-minp",
        "--min_problem_idx",
        type=int,
        help="Min index of problems to run", # limit the number of problems to be run. -1 means no such limit
        default=-1,
    )
    
    # args from open sources models

    parser.add_argument('--model_name_or_path', type=str, help='Path to the model')
    parser.add_argument('--saved_model_path', type=str, help='Path to save the model files')
    parser.add_argument('--hf_dir', type=str, help='Path to the huggingface cache directory')
    parser.add_argument('--input_path', type=str, help='Path to the input file')
    parser.add_argument('--user_input', type=str, help='user input for LLM (testing)')
    parser.add_argument('--output_dir', type=str, help='Path to the output directory')
    parser.add_argument('--chain_length', type=int, default=5, help='Number of steps in the Identity Chain')
    parser.add_argument('--seq_length', type=int, default=8192, help='max length of the sequence')#2048
    parser.add_argument('--gen_length', type=int, default=None, help='max length of the generated sequence')
    parser.add_argument('--do_sample', action='store_true', help='whether to do sampling')
    parser.add_argument('--do_test_only', action='store_true', help='whether to run test for model')
    parser.add_argument('--do_save_model', action='store_true', help='whether to save the model files to a specific directory')
    parser.add_argument('--greedy_early_stop', action='store_true', help='whether to stop inference when fixed point')
    #parser.add_argument('--temperature', type=float, default=0, help='temperature for sampling')
    parser.add_argument('--top_k', type=int, default=0, help='top k for sampling')
    parser.add_argument('--top_p', type=float, default=1, help='top p for sampling')
    parser.add_argument('--num_return_sequences', type=int, default=1, help='number of return sequences')
    parser.add_argument('--num_beams', type=int, default=1, help='number of beams for beam search')
    parser.add_argument('--use_int8', action='store_true', help='whether to use int8 quantization')
    parser.add_argument('--use_fp16', action='store_true', help='whether to use fp16 precision')
    parser.add_argument('--pass_only', action='store_true', help='whether to only pass the input to the next step')
    parser.add_argument('--mask_func_name', action='store_true', help='whether to mask the function name')
    parser.add_argument('--bootstrap_method', type=str, default='problem', help='method to bootstrap the chain')
    parser.add_argument('--resume_task_bs', type=int, default=0, help='task to resume at when bootstrapping')
    parser.add_argument('--resume_task_run', type=int, default=0, help='task to resume at')
    parser.add_argument('--skip_bootstrap', action='store_true', help='whether to skip the bootstrap stage')
    parser.add_argument('--version', type=str, default='v1', help='version of the identity chain')

    args = parser.parse_args()
    model = None
    tokenizer = None
    device = "cuda" if torch.cuda.is_available() else "cpu"
    print('device: ', device)
    if ('Llama' in args.model 
        or args.model.startswith('starcoder')
        or args.model.startswith('deepseek')
        or args.model.startswith('CodeQwen')
        ) and args.log_phase_output != 2:
        # set huggingface cache directory
        HF_HOME = args.hf_dir
        offload_folder = "offload_folder"
        print("Loading model...")
        # if specified, use int8 quantization
        if args.do_save_model:
            model = AutoModelForCausalLM.from_pretrained(
                args.model_name_or_path      
            )
        elif args.use_int8:
            print("**********************************")
            print("**** Using 8-bit quantization ****")
            print("**********************************")
            model = AutoModelForCausalLM.from_pretrained(
                args.model_name_or_path,
                load_in_8bit=True,
                device_map="auto",
                cache_dir=HF_HOME,
                offload_folder=offload_folder,     
            )
        # if specified, use fp16 precision
        elif args.use_fp16:
            print("**********************************")
            print("****** Using fp16 precision ******")
            print("**********************************")
            model = AutoModelForCausalLM.from_pretrained(
                args.model_name_or_path,
                device_map="auto",
                torch_dtype=torch.float16,
                cache_dir=HF_HOME,
                offload_folder=offload_folder,     
            )
        # otherwise, use default precision
        else:
            model = AutoModelForCausalLM.from_pretrained(
                args.model_name_or_path,
                device_map="auto",
                cache_dir=HF_HOME,
                offload_folder=offload_folder,            
            )
        
        # If you want to use multiple GPUs
        #if torch.cuda.device_count() > 1:
        #    model = torch.nn.DataParallel(model)
        print('model device: ', model.device)

        # configure tokenizer
        if (args.model.startswith('Meta-Llama')
            or args.model.startswith('deepseek')
            or args.model.startswith('CodeQwen')):
            tokenizer = AutoTokenizer.from_pretrained(
                args.model_name_or_path,
                trust_remote_code=True,
            )
        else:
            tokenizer = AutoTokenizer.from_pretrained(
                args.model_name_or_path,
                model_max_length=args.seq_length,
                # Bug: A decoder-only architecture is being used, but right-padding was detected! For correct generation results, please set `padding_side='left'` when initializing the tokenizer.
                # setting padding_side='left' doesn't fix the issue.
                padding_side="right",
                use_fast=False,
                trust_remote_code=True,
                cache_dir=HF_HOME,
                offload_folder=offload_folder,
            )
    
    if args.do_test_only:
        test_codellama(tokenizer, model, args.user_input, args.seq_length)
    elif args.do_save_model:
        tokenizer.save_pretrained(args.saved_model_path)
        model.save_pretrained(args.saved_model_path)
    elif args.dataset.startswith('HumanEval'):
        HumanEval_experiment(args.dataset, './Benchmark/HumanEvalComm.jsonl', args.option, args.model, args.topn, args.temperature, args, model, tokenizer)