Spaces:
Sleeping
Sleeping
import time | |
from openai import OpenAI | |
from collections import Counter | |
from loguru import logger | |
import json | |
import os | |
from tqdm import tqdm | |
import sys | |
import io | |
import traceback | |
import ast | |
import time | |
import re | |
from groq import Groq | |
from utils import split_nested_functions, get_dependency_graph_str, evaluate, parse_json_response, extract_code_blocks, extract_functions, extract_function, create_dependency_graph, topological_sort, merge_changes_to_parents, evaluate_simple, parse_transcoder_problem_content | |
from testcase_utils import get_parameter_names, parse_tests | |
# CUDA_VISIBLE_DEVICES=0 python -m vllm.entrypoints.openai.api_server --model deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct --trust-remote-code --dtype auto --api-key token-abc123s --port 18889 | |
# CUDA_VISIBLE_DEVICES=0 python -m vllm.entrypoints.openai.api_server --model TechxGenus/Codestral-22B-v0.1-GPTQ --dtype auto --api-key token-abc123s --port 18890 --trust-remote-code --chat-template helper/codestral_template.jinja | |
# CUDA_VISIBLE_DEVICES=0 python -m vllm.entrypoints.openai.api_server --model Qwen/CodeQwen1.5-7B-Chat --dtype auto --api-key token-abc123s --port 18892 --trust-remote-code | |
MODEL = "deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct" | |
# MODEL = "TechxGenus/Codestral-22B-v0.1-GPTQ" | |
# MODEL = "Qwen/CodeQwen1.5-7B-Chat" | |
client = OpenAI( | |
base_url="http://localhost:18889/v1", | |
api_key="token-abc123s", | |
) | |
# dscoder 18889 | |
# codestral 18890 | |
# codeqwen 18892 | |
# Hyperparameters | |
MAX_VLLM_RETRIES = 10 # maximum number of retries for the VLLM call | |
MAX_PARSE_RETRIES = 3 # we sometimes fail to parse the code/test cases from the response, so we retry | |
MAX_DEBUG_RETRIES = 1 # 3 seems to be slightly better than 1, but the difference is not significant | |
RETRY_DELAY = 0 # seconds | |
REPEAT_CONVERT_HIERARCHICAL_NUM = 1 # seems unimportant | |
REPEAT_TEST_CASE_GENERATION_NUM = 1 # 1 seems to be better than 3 | |
MAX_OUTER_RETRY = 10 # maximum number of retries for the entire debugging process | |
CONTINUOUS_RETRY = False # whether to retry from the last fixed code, else retry from the original buggy code. it seems better to set this to False | |
TEMPERATURE = 0.8 # 0.8 better than 1.0 better than 0.2 | |
MINP = 0.05 | |
# for collecting stats | |
TOTAL_PROMPT_TOKENS = 0 | |
TOTAL_COMPLETION_TOKENS = 0 | |
TOTAL_MG_DEBUG_CALLS = 0 | |
TOTAL_DEBUG_FUNCTION_CALLS = 0 | |
TOTAL_GENERATE_TEST_CASES_CALLS = 0 | |
TOTAL_CONVERT_HIERARCHICAL_CALLS = 0 | |
def get_completion_with_retry(messages, model=MODEL, MAX_VLLM_RETRIES=MAX_VLLM_RETRIES): | |
global TOTAL_PROMPT_TOKENS, TOTAL_COMPLETION_TOKENS | |
for attempt in range(MAX_VLLM_RETRIES): | |
try: | |
logger.info(f"Attempting LLM call (attempt {attempt + 1}/{MAX_VLLM_RETRIES})") | |
logger.info(f"Input messages: {messages[-1]['content']}") | |
chat_completion = client.chat.completions.create( | |
messages=messages, | |
model=MODEL, | |
temperature=TEMPERATURE | |
) | |
response = chat_completion.choices[0].message.content | |
logger.info(f"LLM response: {response}") | |
logger.info("LLM call received") | |
# Update token counts | |
TOTAL_PROMPT_TOKENS += chat_completion.usage.prompt_tokens | |
TOTAL_COMPLETION_TOKENS += chat_completion.usage.completion_tokens | |
return response | |
except Exception as e: | |
logger.error(f"LLM call failed: {str(e)}") | |
if attempt < MAX_VLLM_RETRIES - 1: | |
logger.info(f"Retrying in {RETRY_DELAY} seconds...") | |
time.sleep(RETRY_DELAY) | |
else: | |
logger.error("Max retries reached. Giving up.") | |
raise | |
def generate_test_cases(full_code, gold_test_cases, function_name, MAX_PARSE_RETRIES=MAX_PARSE_RETRIES): | |
global TOTAL_GENERATE_TEST_CASES_CALLS | |
TOTAL_GENERATE_TEST_CASES_CALLS += 1 | |
logger.info(f"Generating test cases for function: {function_name}") | |
prompt = f""" | |
Analyze the following Python code and focus on the function named `{function_name}`. | |
Generate the same number of test cases for this specific function based on the provided gold test cases for the main function. | |
Ensure that the generated test cases are consistent with the behavior expected in the gold test cases. | |
### Full Code | |
{full_code} | |
### Gold Test Cases for the main function | |
{json.dumps(gold_test_cases, indent=2)} | |
### Function to generate test cases for: `{function_name}` | |
### Output format: | |
**Test Case 1:** | |
Input: ... | |
Analysis: ... | |
Expected Output: ... | |
**Test Case 2:** | |
... | |
**All Test Cases:** | |
```json | |
{{ | |
"test_cases": [ | |
{{"input": {{'var1': 'value1', 'var2': 'value2'}}, "expected_output": "expected_output"}}, | |
... | |
] | |
}} | |
``` | |
### Hint | |
Analyze how the `{function_name}` function is used within the main function and how it contributes to the expected outputs in the gold test cases. Generate test cases that reflect this behavior. For each test case, you should analyze step-by-step based on both the input and the expected output of the main function, and then provide the corresponding input and expected output for the `{function_name}` function. | |
""" | |
messages = [ | |
{'role': 'system', 'content': "You are an AI assistant specialized in analyzing Python functions and generating test cases."}, | |
{'role': 'user', 'content': prompt}, | |
] | |
if "starcoder2" in MODEL.lower(): | |
# remove the system message | |
messages = messages[1:] | |
all_test_cases = [] | |
for _ in range(REPEAT_TEST_CASE_GENERATION_NUM): | |
response = get_completion_with_retry(messages) | |
parsed_response = parse_json_response(response) | |
if parsed_response and 'test_cases' in parsed_response: | |
all_test_cases.extend(parsed_response['test_cases']) | |
if not all_test_cases: | |
logger.error("Failed to generate any valid test cases") | |
return [] | |
# Perform majority voting on expected outputs | |
logger.info(f"Generated {len(all_test_cases)} test cases for {function_name}: {all_test_cases}") | |
test_case_counter = Counter([json.dumps(tc) for tc in all_test_cases]) | |
# show the frequency of each test case | |
logger.info(f"Test case frequency:") | |
for tc, freq in test_case_counter.items(): | |
logger.info(f"Test case: {tc}, Frequency: {freq}") | |
most_common_test_cases = test_case_counter.most_common(len(gold_test_cases)) | |
final_test_cases = [json.loads(tc[0]) for tc in most_common_test_cases] | |
logger.info(f"Voted {len(final_test_cases)} test cases for {function_name}: {final_test_cases}") | |
return final_test_cases | |
def debug_function(function_code, function_name, test_cases, MAX_PARSE_RETRIES=MAX_PARSE_RETRIES): | |
global TOTAL_DEBUG_FUNCTION_CALLS | |
TOTAL_DEBUG_FUNCTION_CALLS += 1 | |
logger.info(f"Debugging function:\n{function_code}\nWith test cases: {test_cases}") | |
prompt = f""" | |
Debug the following Python function. The function is not passing all test cases. | |
Analyze the code, identify the bug, and provide a fixed version of the function. | |
### Function: | |
{function_code} | |
### Test Cases: | |
{json.dumps(test_cases, indent=2)} | |
### Task: | |
Provide your response in the following format: | |
1. try to work as a python interpreter to execute the code step-by-step. | |
2. identify the change of each variable as you "run" the code line-by-line. | |
3. based on the execution trace, try to identify the bug | |
4. provide the final fixed code in a Python code block (```python ... ```) | |
Make sure to include the entire function, including the function signature. And you will be rewarded to simulate the code execution in your mind and provide step-by-step trace of the code execution. | |
""" | |
messages = [ | |
{'role': 'system', 'content': 'You are an AI assistant helping to debug Python functions.'}, | |
{'role': 'user', 'content': prompt}, | |
] | |
if "starcoder2" in MODEL.lower(): | |
# remove the system message | |
messages = messages[1:] | |
for attempt in range(MAX_PARSE_RETRIES): | |
try: | |
response = get_completion_with_retry(messages) | |
code_blocks = extract_code_blocks(response) | |
if code_blocks: | |
# search in a reverse order to find the final fixed function | |
for block in code_blocks[::-1]: | |
fixed_function = extract_function(block, function_name) | |
if fixed_function: | |
analysis = response.split("```python")[0].strip() | |
logger.info("Generated debug analysis and fixed code") | |
return analysis, fixed_function | |
raise ValueError("No valid fixed function found in the response") | |
except ValueError as e: | |
logger.error(f"Failed to extract fixed function (attempt {attempt + 1}/{MAX_PARSE_RETRIES}): {str(e)}") | |
if attempt < MAX_PARSE_RETRIES - 1: | |
logger.info(f"Retrying in {RETRY_DELAY} seconds...") | |
time.sleep(RETRY_DELAY) | |
else: | |
logger.error("Max retries reached. Returning None for analysis and fixed code.") | |
return None, None | |
def convert_to_hierarchical(code, include_example=False): | |
global TOTAL_CONVERT_HIERARCHICAL_CALLS | |
TOTAL_CONVERT_HIERARCHICAL_CALLS += 1 | |
logger.info("Converting code to hierarchical structure") | |
example = """ | |
### Example of tree-style hierarchical structure: | |
```python | |
def main_function(input): | |
preprocessed_data = preprocess(input) | |
result = process(preprocessed_data) | |
return result | |
def preprocess(data): | |
cleaned_data = clean_data(data) | |
normalized_data = normalize_data(cleaned_data) | |
return normalized_data | |
def clean_data(data): | |
# Implementation of data cleaning | |
pass | |
def normalize_data(data): | |
# Implementation of data normalization | |
pass | |
def process(data): | |
feature_vector = extract_features(data) | |
result = classify(feature_vector) | |
return result | |
def extract_features(data): | |
# Implementation of feature extraction | |
pass | |
def classify(feature_vector): | |
# Implementation of classification | |
pass | |
``` | |
""" if include_example else "" | |
prompt = f""" | |
Convert the following Python code into a tree-style hierarchical structure with multiple levels of sub-functions. | |
Each significant step or logical block should be its own function, and functions can call other sub-functions. | |
Ensure that the main function calls these sub-functions in the correct order, creating a tree-like structure. | |
### Original Code: | |
{code} | |
{example} | |
### Instructions: | |
Please first analyze the codes step by step, and then provide the converted code in a Python code block (```python ... ```). When providing the final converted code, make sure to include all the functions in a flattened format, where each function is defined separately. | |
""" | |
messages = [ | |
{'role': 'system', 'content': 'You are an AI assistant specialized in refactoring Python code into a tree-style hierarchical structure.'}, | |
{'role': 'user', 'content': prompt}, | |
] | |
if "starcoder2" in MODEL.lower(): | |
# remove the system message | |
messages = messages[1:] | |
best_conversion = None | |
max_subfunctions = 0 | |
for _ in range(REPEAT_CONVERT_HIERARCHICAL_NUM): | |
response = get_completion_with_retry(messages) | |
code_blocks = extract_code_blocks(response) | |
if code_blocks: | |
converted_code = code_blocks[0] | |
subfunctions = len(extract_functions(converted_code)) - 1 # Subtract 1 to exclude the main function | |
if subfunctions > max_subfunctions: | |
max_subfunctions = subfunctions | |
best_conversion = converted_code | |
if best_conversion: | |
logger.info(f"Converted code to tree-style hierarchical structure with {max_subfunctions} sub-functions") | |
# split nested functions | |
best_conversion = split_nested_functions(best_conversion) | |
return best_conversion | |
else: | |
logger.error("Failed to convert code to tree-style hierarchical structure") | |
code = split_nested_functions(code) | |
return code | |
def mg_debug_demo(full_code, gold_test_cases, max_debug_attempts=MAX_DEBUG_RETRIES, given_model=MODEL, given_client=client): | |
global MODEL, client | |
MODEL = given_model | |
client = given_client | |
return mg_debug(full_code, gold_test_cases, max_debug_attempts) | |
def mg_debug(full_code, gold_test_cases, max_debug_attempts=MAX_DEBUG_RETRIES): | |
global TOTAL_MG_DEBUG_CALLS | |
TOTAL_MG_DEBUG_CALLS += 1 | |
logger.info("Starting main debugging process") | |
convert_hierarchical_attempts = 0 | |
while convert_hierarchical_attempts < MAX_PARSE_RETRIES: | |
try: | |
# Convert to tree-style hierarchical structure | |
hierarchical_code = convert_to_hierarchical(full_code, include_example=False) | |
logger.info(f"Converted code to tree-style hierarchical structure:\n{hierarchical_code}") | |
functions = extract_functions(hierarchical_code) | |
# Create a dependency graph | |
dependency_graph = create_dependency_graph(functions) | |
logger.info(f"Dependency graph:\n{get_dependency_graph_str(dependency_graph)}") | |
break | |
except Exception as e: | |
logger.error(f"Failed to convert code to hierarchical structure (attempt {convert_hierarchical_attempts + 1}/{MAX_PARSE_RETRIES}): {str(e)}") | |
convert_hierarchical_attempts += 1 | |
# retry | |
# Sort functions based on their dependencies (bottom-up) | |
sorted_functions = topological_sort(dependency_graph) | |
logger.info(f"Sorted functions: {sorted_functions}") | |
for func_name in sorted_functions: | |
logger.info(f"Processing function: {func_name}") | |
func_code = functions[func_name] | |
test_cases = generate_test_cases(hierarchical_code, gold_test_cases, func_name) | |
fixed_code = func_code | |
for debug_attempt in range(max_debug_attempts): | |
all_tests_pass = True | |
for test_case in test_cases: | |
passed, result = evaluate(hierarchical_code, func_name, test_case) | |
if not passed: | |
all_tests_pass = False | |
break | |
if all_tests_pass: | |
logger.info(f"All tests passed for function: {func_name}") | |
break | |
logger.info(f"Debugging function: {func_name} (Attempt {debug_attempt + 1}/{max_debug_attempts})") | |
analysis, new_fixed_code = debug_function(fixed_code, func_name, test_cases) | |
if new_fixed_code: | |
logger.info(f"New fixed code for {func_name}:\n{new_fixed_code}") | |
fixed_code = new_fixed_code | |
functions[func_name] = fixed_code | |
logger.info(f"Merging {func_name} changes") | |
hierarchical_code = merge_changes_to_parents(func_name, dependency_graph, functions) | |
logger.info(f"Code after merging updates in {func_name}:\n{hierarchical_code}") | |
else: | |
logger.warning(f"Failed to fix {func_name}. Keeping previous implementation.") | |
break | |
if not all_tests_pass: | |
logger.warning(f"Could not fix {func_name} after {max_debug_attempts} attempts. Keeping original implementation.") | |
# Reconstruct the full code with fixed functions | |
fixed_full_code = "\n\n".join(functions.values()) | |
logger.info("Debugging process completed. Reconstructed full code.") | |
return fixed_full_code | |
def test(): | |
buggy_code = ''' | |
def make_palindrome(string: str) -> str: | |
""" Find the shortest palindrome that begins with the supplied string. """ | |
def is_palindrome(s: str) -> bool: | |
return s == s[::-1] | |
suffix_start = 0 | |
for i in range(len(string)): | |
if is_palindrome(string[i:]): | |
suffix_start = i | |
return string + string[:suffix_start][::-1] | |
'''.strip() | |
gold_test_cases = [ | |
{'input': 'cat', 'expected_output': 'catac'}, | |
{'input': 'cata', 'expected_output': 'catac'}, | |
{'input': '', 'expected_output': ''} | |
] | |
entry_point = 'make_palindrome' | |
fixed_code = mg_debug(buggy_code, gold_test_cases) | |
logger.info(f"Fixed code:\n{fixed_code}") | |
logger.info("============= Final evaluation with private test cases =============") | |
# evaluate the final codes with private testcases | |
all_tests = gold_test_cases | |
for testcase in all_tests: | |
result, testcase = evaluate(fixed_code, entry_point, testcase) | |
logger.info(f"Passed: {result}, Test case: {testcase}") | |
def debug_humaneval(input_seeds: str, max_examples: int = None, output_folder: str = None): | |
fixed_problems = 0 | |
total_unsolved = 0 | |
with open(input_seeds, "r") as f: | |
seeds = f.readlines() | |
unsolved_seeds = [] | |
# filter those problems that are not solved | |
for i in range(len(seeds)): | |
problem = json.loads(seeds[i]) | |
if not problem["is_solved"]: | |
unsolved_seeds.append(problem) | |
unsolved_seeds = [seed for seed in unsolved_seeds if seed['task_id'] == 'HumanEval/38'] | |
# resume_path = "output_data/mbpp/seed/codeqwen/20240909-025624/CodeQwen1.5-7B-Chat_debugging_seeds_from_codeqwen.jsonl" | |
# # only load "debugged": false problems from the resume path | |
# with open(resume_path, "r") as f: | |
# resume_problems = f.readlines() | |
# resume_unsolved_task_ids = [json.loads(problem)["task_id"] for problem in resume_problems if not json.loads(problem)["debugged"]] | |
# unsolved_seeds = [problem for problem in unsolved_seeds if problem["task_id"] in resume_unsolved_task_ids] | |
# logger.info(f"Resuming from {resume_path}, {len(unsolved_seeds)} problems to debug") | |
# filter the unsolved problems that are not in the resume path | |
# parse transcoder problems | |
if "transcoder" in input_seeds.lower(): | |
logger.info(f"Parsing the problem content for transcoder problems") | |
unsolved_seeds = [parse_transcoder_problem_content(problem) for problem in tqdm(unsolved_seeds)] | |
total_unsolved = len(unsolved_seeds) | |
logger.info(f"Debugging {total_unsolved} unsolved problems") | |
if max_examples is not None: | |
unsolved_seeds = unsolved_seeds[:max_examples] | |
logger.info(f"Filtering to {max_examples} examples") | |
for problem in tqdm(unsolved_seeds, ncols=100): | |
model_to_be_fixed = input_seeds.split("/")[-2] | |
model_name = MODEL.split("/")[-1] | |
with open(f"{output_folder}/{model_name}_debugging_seeds_from_{model_to_be_fixed}.jsonl", "w+") as f: | |
for seed in unsolved_seeds: | |
f.write(json.dumps(seed) + "\n") | |
logger.info(f"Processing unsolved problem: {problem['task_id']}") | |
logger.info(f"Problem: {problem}") | |
logger.info(f"Problem Raw Prompt: \n{problem['prompt']}") | |
try: | |
buggy_code = problem["solution"] | |
entry_point = problem["entry_point"] | |
try: | |
parameter_names = get_parameter_names(problem["prompt"], entry_point) | |
except: | |
parameter_names = get_parameter_names(problem["solution"], entry_point) | |
logger.info(f"Extracted parameter names: {parameter_names}") | |
# in order to save time, we extract the first 3 given tests for transcoder | |
if "transcoder" in problem["task_id"].lower(): | |
logger.info(f"Extracted {len(problem['given_tests'])} given tests, only using the first 3 samples") | |
problem["given_tests"] = problem["given_tests"][:3] | |
gold_tests_raw = "\n".join(problem["given_tests"]).replace(entry_point, "candidate") | |
gold_tests = parse_tests(gold_tests_raw, parameter_names, entry_point)["test_cases"] | |
logger.info(f"Extracted gold test cases: {gold_tests}") | |
problem['fixed_codes'] = [] # Initialize list to store fixed codes for each retry | |
problem['mg_debug_retries'] = 0 # Initialize retry counter | |
for outer_retry in range(MAX_OUTER_RETRY): | |
fixed_code = mg_debug(buggy_code, gold_tests) | |
problem['fixed_codes'].append(fixed_code) | |
problem['mg_debug_retries'] += 1 | |
all_passed = evaluate_simple(fixed_code, entry_point, problem["test"]) | |
if all_passed: | |
fixed_problems += 1 | |
problem['debugged'] = True | |
logger.info(f"Successfully fixed problem: {problem['task_id']} on retry {outer_retry + 1}") | |
break | |
else: | |
logger.info(f"Failed to fix problem: {problem['task_id']} on retry {outer_retry + 1}") | |
if outer_retry < MAX_OUTER_RETRY - 1: | |
if CONTINUOUS_RETRY: | |
buggy_code = fixed_code # Use the last fixed code as the new buggy code for the next retry | |
else: | |
# Use the original buggy code for the next retry | |
buggy_code = problem["solution"] | |
if not all_passed: | |
problem['debugged'] = False | |
logger.info(f"Failed to fix problem: {problem['task_id']} after {MAX_OUTER_RETRY} retries") | |
except Exception as e: | |
logger.error(f"Error occurred while processing problem: {problem['task_id']}") | |
logger.error(traceback.format_exc()) | |
problem['fixed_codes'] = [] | |
problem['debugged'] = False | |
problem['mg_debug_retries'] = 0 | |
# Add statistics to the log | |
logger.info("=== Statistics ===") | |
logger.info(f"Total prompt tokens: {TOTAL_PROMPT_TOKENS}") | |
logger.info(f"Total completion tokens: {TOTAL_COMPLETION_TOKENS}") | |
logger.info(f"Total mg_debug calls: {TOTAL_MG_DEBUG_CALLS}") | |
logger.info(f"Total debug_function calls: {TOTAL_DEBUG_FUNCTION_CALLS}") | |
logger.info(f"Total generate_test_cases calls: {TOTAL_GENERATE_TEST_CASES_CALLS}") | |
logger.info(f"Total convert_hierarchical calls: {TOTAL_CONVERT_HIERARCHICAL_CALLS}") | |
# Compute and log average statistics | |
avg_prompt_tokens = TOTAL_PROMPT_TOKENS / TOTAL_MG_DEBUG_CALLS if TOTAL_MG_DEBUG_CALLS > 0 else 0 | |
avg_completion_tokens = TOTAL_COMPLETION_TOKENS / TOTAL_MG_DEBUG_CALLS if TOTAL_MG_DEBUG_CALLS > 0 else 0 | |
avg_debug_function_calls = TOTAL_DEBUG_FUNCTION_CALLS / TOTAL_MG_DEBUG_CALLS if TOTAL_MG_DEBUG_CALLS > 0 else 0 | |
avg_generate_test_cases_calls = TOTAL_GENERATE_TEST_CASES_CALLS / TOTAL_MG_DEBUG_CALLS if TOTAL_MG_DEBUG_CALLS > 0 else 0 | |
logger.info(f"Average prompt tokens per mg_debug call: {avg_prompt_tokens:.2f}") | |
logger.info(f"Average completion tokens per mg_debug call: {avg_completion_tokens:.2f}") | |
logger.info(f"Average debug_function calls per mg_debug call: {avg_debug_function_calls:.2f}") | |
logger.info(f"Average generate_test_cases calls per mg_debug call: {avg_generate_test_cases_calls:.2f}") | |
# distribution of debug retries that solved the problem | |
debug_retries = [problem['mg_debug_retries'] for problem in unsolved_seeds if problem['debugged']] | |
debug_retries_counter = Counter(debug_retries) | |
logger.info(f"=== Final Results ===") | |
logger.info(f"Total unsolved problems in seeds: {total_unsolved}") | |
logger.info(f"Problems fixed by our method: {fixed_problems}") | |
logger.info(f"Success rate: {fixed_problems / total_unsolved * 100:.2f}%") | |
logger.info(f"Debug retries distribution for solved problems: {debug_retries_counter}") | |
logger.info(f"Total solved problems before: {len(seeds) - total_unsolved}") | |
logger.info(f"Total solved problems after: {len(seeds) - total_unsolved + fixed_problems}") | |
logger.info(f"Previous accuracy: {(len(seeds) - total_unsolved) / len(seeds) * 100:.2f}%") | |
logger.info(f"Final accuracy: {(len(seeds) - total_unsolved + fixed_problems) / len(seeds) * 100:.2f}%") | |
# save the final results to a file | |
model_to_be_fixed = input_seeds.split("/")[-2] | |
model_name = MODEL.split("/")[-1] | |
with open(f"{output_folder}/{model_name}_debugging_seeds_from_{model_to_be_fixed}.jsonl", "w+") as f: | |
for seed in unsolved_seeds: | |
f.write(json.dumps(seed) + "\n") | |
with open(f"{output_folder}/statistics.json", "w") as f: | |
stats = { | |
"total_prompt_tokens": TOTAL_PROMPT_TOKENS, | |
"total_completion_tokens": TOTAL_COMPLETION_TOKENS, | |
"total_mg_debug_calls": TOTAL_MG_DEBUG_CALLS, | |
"total_debug_function_calls": TOTAL_DEBUG_FUNCTION_CALLS, | |
"total_generate_test_cases_calls": TOTAL_GENERATE_TEST_CASES_CALLS, | |
"total_convert_hierarchical_calls": TOTAL_CONVERT_HIERARCHICAL_CALLS, | |
"avg_prompt_tokens_per_mg_debug": avg_prompt_tokens, | |
"avg_completion_tokens_per_mg_debug": avg_completion_tokens, | |
"avg_debug_function_calls_per_mg_debug": avg_debug_function_calls, | |
"avg_generate_test_cases_calls_per_mg_debug": avg_generate_test_cases_calls | |
} | |
json.dump(stats, f, indent=2) | |
scores = { | |
"total_unsolved": total_unsolved, | |
"fixed_problems": fixed_problems, | |
"success_rate": fixed_problems / total_unsolved * 100, | |
"debug_retries_distribution": debug_retries_counter, | |
"total_solved_problems_before": len(seeds) - total_unsolved, | |
"total_solved_problems_after": len(seeds) - total_unsolved + fixed_problems, | |
"previous_accuracy": (len(seeds) - total_unsolved) / len(seeds) * 100, | |
"final_accuracy": (len(seeds) - total_unsolved + fixed_problems) / len(seeds) * 100 | |
} | |
json.dump(scores, f, indent=2) | |
if __name__ == "__main__": | |
input_seeds = "input_data/humaneval/seed/reflexion/seed.jsonl" | |
# input_seeds = "input_data/humaneval/seed/codestral/seed.jsonl" | |
# input_seeds = "input_data/humaneval/seed/codeqwen/seed.jsonl" | |
# input_seeds = "input_data/mbpp/seed/codestral/seed.jsonl" | |
# input_seeds = "input_data/mbpp/seed/deepseekcoder/seed.jsonl" | |
# input_seeds = "input_data/mbpp/seed/codeqwen/seed.jsonl" | |
# input_seeds = "input_data/humanevalfix/seeds.jsonl" | |
seed_stamp = input_seeds.split("input_data/")[-1].replace("/seed.jsonl", "") | |
timestamp = time.strftime("%Y%m%d-%H%M%S") | |
output_folder = f"./output_data/{seed_stamp}/{timestamp}" | |
os.makedirs(output_folder, exist_ok=True) | |
# Configure logger | |
logger.remove() | |
logger.add(sys.stderr, format="<green>{time:YYYY-MM-DD HH:mm:ss}</green> | <level>{level: <8}</level> | <cyan>{name}</cyan>:<cyan>{function}</cyan>:<cyan>{line}</cyan> - <level>{message}</level>") | |
logger.add(f"{output_folder}/all_info.log", format="<green>{time:YYYY-MM-DD HH:mm:ss}</green> | <level>{level: <8}</level> | <cyan>{name}</cyan>:<cyan>{function}</cyan>:<cyan>{line}</cyan> - <level>{message}</level>") | |
# save all important params in a file | |
with open(f"{output_folder}/params.json", "w+") as f: | |
info = { | |
"input_seeds": input_seeds, | |
"output_folder": output_folder, | |
"MAX_DEBUG_RETRIES": MAX_DEBUG_RETRIES, | |
"MAX_PARSE_RETRIES": MAX_PARSE_RETRIES, | |
"RETRY_DELAY": RETRY_DELAY, | |
"REPEAT_CONVERT_HIERARCHICAL_NUM": REPEAT_CONVERT_HIERARCHICAL_NUM, | |
"REPEAT_TEST_CASE_GENERATION_NUM": REPEAT_TEST_CASE_GENERATION_NUM, | |
"MODEL": MODEL, | |
"MAX_OUTER_RETRY": MAX_OUTER_RETRY, | |
"CONTINUOUS_RETRY": CONTINUOUS_RETRY, | |
"TEMPERATURE": TEMPERATURE, | |
"MAX_VLLM_RETRIES": MAX_VLLM_RETRIES, | |
"MINP": MINP | |
} | |
f.write(json.dumps(info, indent=2)) | |
debug_humaneval(input_seeds, max_examples=None, output_folder=output_folder) | |