generate-readme-eval / _script_for_gen.py
codelion's picture
Upload 2 files
058ded2 verified
raw
history blame
17.1 kB
import os
import argparse
import json
import numpy as np
from tqdm import tqdm
import nltk
from nltk.translate.bleu_score import sentence_bleu, SmoothingFunction
from rouge import Rouge
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.metrics.pairwise import cosine_similarity
import re
from textstat import flesch_reading_ease
from datasets import load_dataset
import openai
import time
from datetime import datetime
import google.generativeai as genai
import traceback
SLEEP_INTERVAL = 30
nltk.download('punkt', quiet=True)
nltk.download('averaged_perceptron_tagger', quiet=True)
def create_client(model_name, base_url):
if model_name.lower().startswith('gemini'):
api_key = os.getenv("GOOGLE_API_KEY")
if not api_key:
raise ValueError("GOOGLE_API_KEY environment variable is not set")
genai.configure(api_key=api_key)
return 'gemini'
else:
api_key = os.getenv("OPENAI_API_KEY")
if not api_key:
raise ValueError("OPENAI_API_KEY environment variable is not set")
return openai.OpenAI(api_key=api_key) if base_url is None else openai.OpenAI(api_key=api_key, base_url=base_url)
def preprocess(text):
return nltk.word_tokenize(text.lower())
def calculate_bleu(reference, candidate):
reference_tokens = preprocess(reference)
candidate_tokens = preprocess(candidate)
smoothie = SmoothingFunction().method1
return sentence_bleu([reference_tokens], candidate_tokens, smoothing_function=smoothie)
def calculate_rouge(reference, candidate):
rouge = Rouge()
scores = rouge.get_scores(candidate, reference)
return {
'rouge-1': scores[0]['rouge-1']['f'],
'rouge-2': scores[0]['rouge-2']['f'],
'rouge-l': scores[0]['rouge-l']['f']
}
def calculate_cosine_similarity(reference, candidate):
vectorizer = TfidfVectorizer()
tfidf_matrix = vectorizer.fit_transform([reference, candidate])
return cosine_similarity(tfidf_matrix[0:1], tfidf_matrix[1:2])[0][0]
def extract_sections(readme):
sections = []
current_section = ""
for line in readme.split('\n'):
if line.strip().startswith('#'):
if current_section:
sections.append(current_section.strip())
current_section = line + "\n"
else:
current_section += line + "\n"
if current_section:
sections.append(current_section.strip())
return sections
def calculate_structural_similarity(reference, candidate):
ref_sections = extract_sections(reference)
cand_sections = extract_sections(candidate)
# Calculate section difference
max_sections = max(len(ref_sections), len(cand_sections))
section_diff = abs(len(ref_sections) - len(cand_sections))
section_similarity = 1 - (section_diff / max_sections) if max_sections > 0 else 0
# Calculate title similarity
ref_titles = [s.split('\n')[0] for s in ref_sections]
cand_titles = [s.split('\n')[0] for s in cand_sections]
title_similarity = len(set(ref_titles) & set(cand_titles)) / max(len(ref_titles), len(cand_titles)) if ref_titles or cand_titles else 0
# Combine section and title similarity
structural_similarity = (section_similarity + title_similarity) / 2
return structural_similarity
def information_retrieval_score(readme):
key_sections = ['installation', 'usage', 'api', 'example', 'license']
found_sections = sum(1 for section in key_sections if section in readme.lower())
return found_sections / len(key_sections)
def code_readme_consistency(repo_content, readme):
code_elements = set(re.findall(r'def\s+(\w+)', repo_content) +
re.findall(r'class\s+(\w+)', repo_content))
mentioned_elements = sum(1 for element in code_elements if element in readme)
return mentioned_elements / len(code_elements) if code_elements else 0
def calculate_readability(text):
return flesch_reading_ease(text) / 100
def evaluate_readme(reference_readme, generated_readme, repo_content):
bleu_score = calculate_bleu(reference_readme, generated_readme)
rouge_scores = calculate_rouge(reference_readme, generated_readme)
cosine_sim = calculate_cosine_similarity(reference_readme, generated_readme)
structural_sim = calculate_structural_similarity(reference_readme, generated_readme)
info_retrieval = information_retrieval_score(generated_readme)
code_consistency = code_readme_consistency(repo_content, generated_readme)
readability = calculate_readability(generated_readme)
weights = {
'bleu': 0.1,
'rouge-1': 0.033,
'rouge-2': 0.033,
'rouge-l': 0.034,
'cosine_similarity': 0.1,
'structural_similarity': 0.1,
'information_retrieval': 0.2,
'code_consistency': 0.2,
'readability': 0.2
}
weighted_score = (
weights['bleu'] * bleu_score +
weights['rouge-1'] * rouge_scores['rouge-1'] +
weights['rouge-2'] * rouge_scores['rouge-2'] +
weights['rouge-l'] * rouge_scores['rouge-l'] +
weights['cosine_similarity'] * cosine_sim +
weights['structural_similarity'] * structural_sim +
weights['information_retrieval'] * info_retrieval +
weights['code_consistency'] * code_consistency +
weights['readability'] * readability
)
return {
'bleu': bleu_score,
'rouge': rouge_scores,
'cosine_similarity': cosine_sim,
'structural_similarity': structural_sim,
'information_retrieval': info_retrieval,
'code_consistency': code_consistency,
'readability': readability,
'weighted_score': weighted_score
}
def generate_readme_openai(repo_content, model, client):
system_prompt = """You are an AI assistant tasked with creating a README.md file for a GitHub repository.
Your response should contain ONLY the content of the README.md file, without any additional explanations or markdown code blocks.
The README should include the following sections:
1. Project Title
2. Description
3. Installation
4. Usage
5. Features
6. Contributing
7. License
Ensure that your response is well-structured, informative, and directly usable as a README.md file."""
user_prompt = f"Here is the content of the repository:\n\n{repo_content}\n\nBased on this content, please generate a README.md file."
response = client.chat.completions.create(
model=model,
messages=[
{"role": "system", "content": system_prompt},
{"role": "user", "content": user_prompt}
]
)
return response.choices[0].message.content
def generate_readme_gemini(repo_content, model):
safe = [
{
"category": "HARM_CATEGORY_HARASSMENT",
"threshold": "BLOCK_NONE",
},
{
"category": "HARM_CATEGORY_HATE_SPEECH",
"threshold": "BLOCK_NONE",
},
{
"category": "HARM_CATEGORY_SEXUALLY_EXPLICIT",
"threshold": "BLOCK_NONE",
},
{
"category": "HARM_CATEGORY_DANGEROUS_CONTENT",
"threshold": "BLOCK_NONE",
},
]
prompt = f"""Create a README.md file for a GitHub repository based on the following repository content.
Your response should contain ONLY the content of the README.md file, without any additional explanations or markdown code blocks.
The README should include the following sections:
1. Project Title
2. Description
3. Installation
4. Usage
5. Features
6. Contributing
7. License
Ensure that your response is well-structured, informative, and directly usable as a README.md file.
Repository content:
{repo_content}
"""
model = genai.GenerativeModel(model,safety_settings=safe)
response = model.generate_content(prompt)
return response.text
def generate_readme(repo_content, model_name, client):
if client == 'gemini':
return generate_readme_gemini(repo_content, model_name)
else:
return generate_readme_openai(repo_content, model_name, client)
def main(args):
dataset = load_dataset("patched-codes/generate-readme-eval")
results = []
if args.generate_fine_tune_jsonl:
output_file = f"fine_tune_data_{datetime.now().strftime('%Y%m%d_%H%M%S')}.jsonl"
generate_fine_tune_jsonl(dataset, output_file)
print(f"Fine-tune JSONL file generated: {output_file}")
return
if args.oracle:
model_name = "oracle"
else:
model_name = args.model
client = create_client(model_name, args.base_url)
for item in tqdm(dataset['test'], desc="Processing repos"):
try:
if args.oracle:
# Use the existing README as both reference and generated
generated_readme = item['repo_readme']
elif args.n_shot > 0:
generated_readme = generate_readme_n_shot(item['repo_content'], model_name, client, dataset['train'], args.n_shot)
else:
generated_readme = generate_readme(item['repo_content'], model_name, client)
eval_result = evaluate_readme(item['repo_readme'], generated_readme, item['repo_content'])
eval_result['repo_name'] = item['repo_name']
results.append(eval_result)
except Exception as e:
print(f"Error processing repo {item['repo_name']}: {e}")
continue
if model_name.lower().startswith('gemini'):
time.sleep(SLEEP_INTERVAL)
average_scores = {
'bleu': np.mean([r['bleu'] for r in results]),
'rouge-1': np.mean([r['rouge']['rouge-1'] for r in results]),
'rouge-2': np.mean([r['rouge']['rouge-2'] for r in results]),
'rouge-l': np.mean([r['rouge']['rouge-l'] for r in results]),
'cosine_similarity': np.mean([r['cosine_similarity'] for r in results]),
'structural_similarity': np.mean([r['structural_similarity'] for r in results]),
'information_retrieval': np.mean([r['information_retrieval'] for r in results]),
'code_consistency': np.mean([r['code_consistency'] for r in results]),
'readability': np.mean([r['readability'] for r in results]),
'weighted_score': np.mean([r['weighted_score'] for r in results])
}
# Print results to console
print("\nEvaluation Results:")
for metric, score in average_scores.items():
print(f"{metric}: {score:.4f}")
# Save results to log file
timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
log_filename = f"{model_name}_results_{timestamp}.log"
with open(log_filename, 'w') as log_file:
log_file.write(f"Evaluation Results for model: {model_name}\n")
log_file.write(f"Timestamp: {timestamp}\n\n")
log_file.write("Average Scores:\n")
for metric, score in average_scores.items():
log_file.write(f"{metric}: {score:.4f}\n")
log_file.write(f"\nDetailed Results:\n")
for result in results:
log_file.write(f"\nRepository: {result['repo_name']}\n")
log_file.write("Scores:\n")
log_file.write(f" BLEU: {result['bleu']:.4f}\n")
log_file.write(f" ROUGE-1: {result['rouge']['rouge-1']:.4f}\n")
log_file.write(f" ROUGE-2: {result['rouge']['rouge-2']:.4f}\n")
log_file.write(f" ROUGE-L: {result['rouge']['rouge-l']:.4f}\n")
log_file.write(f" Cosine Similarity: {result['cosine_similarity']:.4f}\n")
log_file.write(f" Structural Similarity: {result['structural_similarity']:.4f}\n")
log_file.write(f" Information Retrieval: {result['information_retrieval']:.4f}\n")
log_file.write(f" Code Consistency: {result['code_consistency']:.4f}\n")
log_file.write(f" Readability: {result['readability']:.4f}\n")
log_file.write(f" Weighted Score: {result['weighted_score']:.4f}\n")
print(f"\nResults saved to {log_filename}")
def generate_fine_tune_jsonl(dataset, output_file):
system_prompt = """You are an AI assistant tasked with creating a README.md file for a GitHub repository.
Your response should contain ONLY the content of the README.md file, without any additional explanations or markdown code blocks.
The README should include the following sections:
1. Project Title
2. Description
3. Installation
4. Usage
5. Features
6. Contributing
7. License
Ensure that your response is well-structured, informative, and directly usable as a README.md file."""
with open(output_file, 'w') as f:
for item in tqdm(dataset['train'], desc="Generating fine-tune data"):
user_prompt = f"Here is the content of the repository:\n\n{item['repo_content']}\n\nBased on this content, please generate a README.md file."
messages = [
{"role": "system", "content": system_prompt},
{"role": "user", "content": user_prompt},
{"role": "assistant", "content": item['repo_readme']}
]
json.dump({"messages": messages}, f)
f.write('\n')
def find_similar_examples(repo_content, train_dataset, n):
vectorizer = TfidfVectorizer()
train_contents = [item['repo_content'] for item in train_dataset]
train_vectors = vectorizer.fit_transform(train_contents)
query_vector = vectorizer.transform([repo_content])
similarities = cosine_similarity(query_vector, train_vectors).flatten()
top_n_indices = similarities.argsort()[-n:][::-1]
return [train_dataset[int(i)] for i in top_n_indices]
def generate_readme_n_shot(repo_content, model_name, client, train_dataset, n_shot):
similar_examples = find_similar_examples(repo_content, train_dataset, n_shot)
system_prompt = """You are an AI assistant tasked with creating a README.md file for a GitHub repository.
Your response should contain ONLY the content of the README.md file, without any additional explanations or markdown code blocks.
The README should include the following sections:
1. Project Title
2. Description
3. Installation
4. Usage
5. Features
6. Contributing
7. License
Ensure that your response is well-structured, informative, and directly usable as a README.md file."""
few_shot_examples = ""
for example in similar_examples:
few_shot_examples += f"Repository content:\n\n{example['repo_content']}\n\n"
few_shot_examples += f"Generated README:\n\n{example['repo_readme']}\n\n---\n\n"
user_prompt = f"""Here are some examples of repository contents and their corresponding README files:
{few_shot_examples}
Now, here is the content of the repository you need to create a README for:
{repo_content}
Based on this content and the examples provided, please generate a README.md file."""
if client == 'gemini':
safe = [
{"category": "HARM_CATEGORY_HARASSMENT", "threshold": "BLOCK_NONE"},
{"category": "HARM_CATEGORY_HATE_SPEECH", "threshold": "BLOCK_NONE"},
{"category": "HARM_CATEGORY_SEXUALLY_EXPLICIT", "threshold": "BLOCK_NONE"},
{"category": "HARM_CATEGORY_DANGEROUS_CONTENT", "threshold": "BLOCK_NONE"},
]
model = genai.GenerativeModel(model_name, safety_settings=safe)
response = model.generate_content(user_prompt)
return response.text
else:
response = client.chat.completions.create(
model=model_name,
messages=[
{"role": "system", "content": system_prompt},
{"role": "user", "content": user_prompt}
]
)
return response.choices[0].message.content
if __name__ == "__main__":
parser = argparse.ArgumentParser(description="Generate and evaluate README files using OpenAI or Gemini API, or compute oracle scores")
parser.add_argument("model", nargs='?', help="Model to use (e.g., 'gpt-4o-mini' for OpenAI or 'gemini-1.5-flash' for Google)")
parser.add_argument("--base_url", help="Optional base URL for OpenAI API", default=None)
parser.add_argument("--oracle", action="store_true", help="Compute oracle scores using existing READMEs")
parser.add_argument("--generate-fine-tune-jsonl", action="store_true", help="Generate a JSONL file for fine-tuning")
parser.add_argument("--n_shot", type=int, default=0, help="Number of examples to use for few-shot learning")
args = parser.parse_args()
if args.generate_fine_tune_jsonl:
if args.oracle or args.model:
parser.error("--generate-fine-tune-jsonl flag cannot be used with --oracle or model specification")
elif args.oracle and args.model:
parser.error("--oracle flag cannot be used with a model specification")
elif not args.oracle and not args.model:
parser.error("Either --oracle flag or a model name must be provided")
main(args)