README.md

metadata

dataset_info:
  - config_name: vllm
    features:
      - name: commit_hash
        dtype: string
      - name: pr_url
        dtype: string
      - name: pr_date
        dtype: string
      - name: timeline_extracted_at
        dtype: string
      - name: analysis_extracted_at
        dtype: string
      - name: models
        list: string
      - name: perf_command
        dtype: string
      - name: has_serving
        dtype: bool
      - name: has_latency
        dtype: bool
      - name: has_throughput
        dtype: bool
      - name: uses_lm_eval
        dtype: bool
      - name: commit_subject
        dtype: string
      - name: commit_message
        dtype: string
      - name: commit_date
        dtype: string
      - name: files_changed
        list: string
      - name: stats
        struct:
          - name: commit_year
            dtype: int64
          - name: num_edited_lines
            dtype: int64
          - name: num_files
            dtype: int64
          - name: num_hunks
            dtype: int64
          - name: num_non_test_edited_lines
            dtype: int64
          - name: num_non_test_files
            dtype: int64
          - name: num_test_files
            dtype: int64
          - name: only_non_test_files
            dtype: int64
          - name: only_test_files
            dtype: int64
      - name: diff_text
        dtype: string
      - name: apis
        list: string
      - name: affected_paths
        list: string
      - name: repo
        dtype: string
      - name: hardware
        dtype: string
      - name: lm_eval_command
        dtype: string
    splits:
      - name: train
        num_bytes: 545364
        num_examples: 39
    download_size: 192570
    dataset_size: 545364
  - config_name: sglang
    features:
      - name: commit_hash
        dtype: string
      - name: pr_url
        dtype: string
      - name: pr_date
        dtype: string
      - name: timeline_extracted_at
        dtype: string
      - name: analysis_extracted_at
        dtype: string
      - name: models
        list: string
      - name: perf_command
        dtype: string
      - name: has_serving
        dtype: bool
      - name: has_latency
        dtype: bool
      - name: has_throughput
        dtype: bool
      - name: uses_lm_eval
        dtype: bool
      - name: commit_subject
        dtype: string
      - name: commit_message
        dtype: string
      - name: commit_date
        dtype: string
      - name: files_changed
        list: string
      - name: stats
        struct:
          - name: commit_year
            dtype: int64
          - name: num_edited_lines
            dtype: int64
          - name: num_files
            dtype: int64
          - name: num_hunks
            dtype: int64
          - name: num_non_test_edited_lines
            dtype: int64
          - name: num_non_test_files
            dtype: int64
          - name: num_test_files
            dtype: int64
          - name: only_non_test_files
            dtype: int64
          - name: only_test_files
            dtype: int64
      - name: diff_text
        dtype: string
      - name: apis
        list: string
      - name: affected_paths
        list: string
      - name: repo
        dtype: string
      - name: hardware
        dtype: string
      - name: lm_eval_command
        dtype: string
    splits:
      - name: train
        num_bytes: 91137
        num_examples: 15
    download_size: 52410
    dataset_size: 91137
configs:
  - config_name: vllm
    data_files:
      - split: train
        path: vllm/train-*
  - config_name: sglang
    data_files:
      - split: train
        path: sglang/train-*
license: apache-2.0
task_categories:
  - text-generation
language:
  - en
size_categories:
  - n<1K
tags:
  - performance-optimization
  - software-engineering
  - benchmark
  - ai-agents
  - code

ISO-Bench Dataset

A curated dataset of real-world software performance optimization commits from vLLM and SGLang, designed for evaluating AI agents on code optimization tasks.

Dataset Summary

Config	Commits	Repository
vllm	39	vLLM (LLM inference engine)
sglang	15	SGLang (LLM serving framework)

Each entry represents a human-authored performance optimization commit with:

• The original commit diff and message
• Performance benchmark commands (perf_command)
• Model configurations for benchmarking
• Hardware requirements
• API surface analysis

Usage

from datasets import load_dataset

# Load vLLM optimization commits
vllm = load_dataset('Lossfunk/ISO-Bench', 'vllm', split='train')

# Load SGLang optimization commits
sglang = load_dataset('Lossfunk/ISO-Bench', 'sglang', split='train')

# Example: inspect a commit
print(vllm[0]['commit_subject'])
print(vllm[0]['perf_command'])
print(vllm[0]['models'])

Schema

Field	Type	Description
commit_hash	string	Short hash of the optimization commit
pr_url	string	URL to the pull request
commit_subject	string	Commit message subject line
commit_message	string	Full commit message
diff_text	string	Unified diff of the optimization
models	list[string]	HuggingFace model IDs used for benchmarking
perf_command	string	Command to run the performance benchmark
has_serving	bool	Whether commit affects serving performance
has_latency	bool	Whether commit affects latency
has_throughput	bool	Whether commit affects throughput
uses_lm_eval	bool	Whether correctness is validated via lm-eval
lm_eval_command	string	lm-eval command for correctness validation
files_changed	list[string]	Files modified in the commit
apis	list[string]	Affected API endpoints/functions
affected_paths	list[string]	Code paths affected by the change
hardware	string	Required hardware (e.g., GPU type)
stats	struct	Commit statistics (lines changed, files, hunks)

How It Works

Each dataset entry captures a real performance optimization made by an expert developer. AI agents are given the codebase at the parent commit (before optimization) and must independently discover and implement a performance improvement. Their patches are then benchmarked against the human expert's solution using wall-clock timing comparisons.

License

Apache 2.0