Inferencebench/vllm-pr-analysis

vLLM PR Test Classification Dataset

🎯 Overview

This dataset contains 98 vLLM project commits with their corresponding Pull Request (PR) timeline data and comprehensive test type classifications. It provides insights into testing patterns in a major LLM serving infrastructure project.

📊 Dataset Description

Purpose

This dataset was created by analyzing vLLM project PR timelines to:

• Identify different types of testing and benchmarking activities
• Understand testing patterns in LLM infrastructure development
• Provide labeled data for ML models to classify test types in software PRs
• Enable research on performance optimization trends in LLM serving

Test Categories

Each commit is classified across four test categories:

Category	Description	Keywords	Prevalence
LM Evaluation	Language model evaluation tests	lm_eval, gsm8k, mmlu, hellaswag, truthfulqa	25.5%
Performance	Performance benchmarking tests	TTFT, throughput, latency, ITL, TPOT, tok/s	81.6%
Serving	Serving functionality tests	vllm serve, API server, frontend, online serving	53.1%
General Test	General testing activities	CI, pytest, unittest, buildkite, fastcheck	96.9%

📈 Dataset Statistics

Overall Distribution

• Total commits: 98
• Multi-category commits: 76 (77.6%)
• Average test types per commit: 2.57

Detailed Keyword Frequency

Top Performance Keywords (80 commits)

• throughput: 241 mentions
• latency: 191 mentions
• profiling: 114 mentions
• TTFT (Time To First Token): 114 mentions
• ITL (Inter-token Latency): 114 mentions
• TPOT (Time Per Output Token): 108 mentions
• optimization: 87 mentions
• tok/s (tokens per second): 66 mentions

Top LM Evaluation Keywords (25 commits)

• gsm8k: 62 mentions
• lm_eval: 33 mentions
• lm-eval: 9 mentions
• mmlu: 3 mentions
• humaneval: 1 mention

Top Serving Keywords (52 commits)

• frontend: 181 mentions
• serving: 74 mentions
• api server: 42 mentions
• vllm serve: 23 mentions
• online serving: 19 mentions

🗂️ Data Schema

{
    'commit_hash': str,           # Git commit SHA-1 hash (40 chars)
    'pr_url': str,                # GitHub PR URL (e.g., https://github.com/vllm-project/vllm/pull/12601)
    'has_lm_eval': bool,          # True if commit contains LM evaluation tests
    'has_performance': bool,       # True if commit contains performance benchmarks
    'has_serving': bool,          # True if commit contains serving tests
    'has_general_test': bool,     # True if commit contains general tests
    'test_details': str,          # Pipe-separated test keywords (e.g., "PERF: ttft, throughput | TEST: ci, pytest")
    'timeline_text': str,         # Full PR timeline text with comments, reviews, and commit messages
    'extracted_at': str           # ISO timestamp when data was extracted
}

💻 Usage Examples

Basic Loading

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("your-username/vllm-pr-test-classification")

# Explore the data
print(f"Total examples: {len(dataset['train'])}")
print(f"Features: {dataset['train'].features}")
print(f"First example: {dataset['train'][0]}")

Filtering Examples

# Get commits with performance benchmarks
perf_commits = dataset['train'].filter(lambda x: x['has_performance'])
print(f"Performance commits: {len(perf_commits)}")

# Get commits with LM evaluation
lm_eval_commits = dataset['train'].filter(lambda x: x['has_lm_eval'])
print(f"LM evaluation commits: {len(lm_eval_commits)}")

# Get commits with multiple test types
multi_test = dataset['train'].filter(
    lambda x: sum([x['has_lm_eval'], x['has_performance'], 
                   x['has_serving'], x['has_general_test']]) >= 3
)
print(f"Commits with 3+ test types: {len(multi_test)}")

Analysis Example

import pandas as pd

# Convert to pandas for analysis
df = dataset['train'].to_pandas()

# Analyze test type combinations
test_combinations = df[['has_lm_eval', 'has_performance', 'has_serving', 'has_general_test']]
combination_counts = test_combinations.value_counts()
print("Most common test combinations:")
print(combination_counts.head())

# Extract performance metrics mentioned
perf_df = df[df['has_performance']]
print(f"\nCommits mentioning specific metrics:")
print(f"TTFT mentions: {perf_df['test_details'].str.contains('TTFT').sum()}")
print(f"Throughput mentions: {perf_df['test_details'].str.contains('throughput', case=False).sum()}")

Text Classification Training

from transformers import AutoTokenizer, AutoModelForSequenceClassification
from transformers import TrainingArguments, Trainer

# Prepare for multi-label classification
def preprocess_function(examples):
    # Create multi-label targets
    labels = []
    for i in range(len(examples['commit_hash'])):
        label = [
            int(examples['has_lm_eval'][i]),
            int(examples['has_performance'][i]),
            int(examples['has_serving'][i]),
            int(examples['has_general_test'][i])
        ]
        labels.append(label)
    
    # Tokenize timeline text
    tokenized = tokenizer(
        examples['timeline_text'],
        truncation=True,
        padding='max_length',
        max_length=512
    )
    tokenized['labels'] = labels
    return tokenized

# Train a classifier to identify test types from PR text
tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")
model = AutoModelForSequenceClassification.from_pretrained(
    "bert-base-uncased",
    num_labels=4,
    problem_type="multi_label_classification"
)

🔍 Sample Data

Example 1: Performance-focused commit

{
  "commit_hash": "fc542144c4477ffec1d3de6fa43e54f8fb5351e8",
  "pr_url": "https://github.com/vllm-project/vllm/pull/12563",
  "has_lm_eval": false,
  "has_performance": true,
  "has_serving": false,
  "has_general_test": true,
  "test_details": "PERF: tok/s, optimization | TEST: CI",
  "timeline_text": "[Guided decoding performance optimization]..."
}

Example 2: Comprehensive testing commit

{
  "commit_hash": "aea94362c9bdd08ed2b346701bdc09d278e85f66",
  "pr_url": "https://github.com/vllm-project/vllm/pull/12287",
  "has_lm_eval": true,
  "has_performance": true,
  "has_serving": true,
  "has_general_test": true,
  "test_details": "LM_EVAL: lm_eval, gsm8k | PERF: TTFT, ITL | SERVING: vllm serve | TEST: test, CI",
  "timeline_text": "[Frontend][V1] Online serving performance improvements..."
}

🛠️ Potential Use Cases

1. Test Type Classification: Train models to automatically classify test types in software PRs
2. Testing Pattern Analysis: Study how different test types correlate in infrastructure projects
3. Performance Optimization Research: Analyze performance testing trends in LLM serving systems
4. CI/CD Insights: Understand continuous integration patterns in ML infrastructure projects
5. Documentation Generation: Generate test documentation from PR timelines
6. Code Review Automation: Build tools to automatically suggest relevant tests based on PR content

📚 Source

This dataset was extracted from the vLLM project GitHub repository PR timelines. vLLM is a high-throughput and memory-efficient inference and serving engine for LLMs.

🔄 Updates

• v1.0.0 (2025-01): Initial release with 98 commits

📜 License

This dataset is released under the MIT License, consistent with the vLLM project's licensing.

📖 Citation

If you use this dataset in your research or applications, please cite:

@dataset{vllm_pr_test_classification_2025,
  title={vLLM PR Test Classification Dataset},
  author={vLLM Community Contributors},
  year={2025},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/your-username/vllm-pr-test-classification},
  note={A dataset of 98 vLLM commits with test type classifications extracted from GitHub PR timelines}
}

🤝 Contributing

If you'd like to contribute to this dataset or report issues:

1. Open an issue on the Hugging Face dataset repository
2. Submit improvements via pull requests
3. Share your use cases and findings

⚠️ Limitations

• Dataset size is limited to 98 commits
• Timeline text may be truncated for very long PR discussions
• Classification is based on keyword matching, which may miss context-dependent references
• Dataset represents a snapshot from specific time period of vLLM development

🙏 Acknowledgments

Thanks to the vLLM project maintainers and contributors for their open-source work that made this dataset possible.