LLMSYS-HPOBench

LLMSYS-HPOBench is an offline benchmark dataset for hyperparameter optimization of real-world LLM systems. It covers inference engines, RAG pipelines, and agent frameworks, with normalized tabular measurements linked to log and hardware artifacts when available.

Project Links

• GitHub repository, benchmark loader, and contribution guide: https://github.com/ideas-labo/llmsys-hpobench
• Paper: https://arxiv.org/abs/2605.08305
• Full data archive on Zenodo: https://zenodo.org/records/20048594

What Is Hosted Here

This Hugging Face dataset repository hosts the dataset card, Croissant metadata, schema documentation, the sample manifest, a small example data package, and a Hugging Face-native version of the full normalized benchmark payload.

The original directory-shaped data package is mirrored on Zenodo for citation-friendly archival downloads. In this Hugging Face repository, the same benchmark rows are exposed as Parquet tables and the many small log/hardware text files are packed into indexed .tar.zst artifact shards.

Dataset Structure

The original benchmark data is organized by category and system:

experiment-data/
|-- Agent/
|   |-- autogpt/
|   `-- openhands/
|-- Engine/
|   |-- SGLang/
|   `-- vLLM/
`-- RAG/
    |-- html_rag/
    |-- LightRAG/
    `-- naiverag/

Each original fidelity directory contains one normalized CSV and optional per-sample artifacts:

{system}/{fidelity_name}/
|-- {fidelity_name}.csv
|-- log_file/
|   `-- log-{ID}.txt
`-- hw_file/
    `-- hw-{ID}.txt

Main CSVs follow this schema:

Column type	Format
Row ID	ID
AI hyperparameters	cfg-ai-{name}
Non-AI hyperparameters	cfg-{name}
Objective metrics	obj-{name}+ or obj-{name}-
Cost metrics	cost-{name}
Hardware artifact	hw-file
Combined log artifact	log-file

See format.md for the full cleaning and naming specification.

The Hugging Face-native layout is:

data/
|-- records/records.parquet
|-- config_long/config_long.parquet
|-- metrics_long/metrics_long.parquet
`-- artifacts_index/artifacts_index.parquet
artifacts/
|-- logs/*.tar.zst
`-- hardware/*.tar.zst

• records.parquet: one row per benchmark sample, with JSON fields for grouped configuration, objective, cost, hardware, and original row values.
• config_long.parquet: one row per configuration value, with is_ai marking AI vs non-AI hyperparameters.
• metrics_long.parquet: one row per objective, cost, or expanded hardware metric.
• artifacts_index.parquet: maps each sample artifact reference to a .tar.zst archive and member path.

Usage

Clone the code repository and use either the Hugging Face-native Parquet/artifact layout or the original Zenodo archive:

git clone https://github.com/ideas-labo/llmsys-hpobench
cd llmsys-hpobench
# For the original loader, extract the Zenodo archive so ./experiment-data exists.
uv run python llmsys_hpobench.py --root experiment-data --system vLLM --budget 3

Python example:

from llmsys_hpobench import Benchmark

benchmark = Benchmark(system="vLLM", root="experiment-data")
config = benchmark.get_config_space().sample(random_state=0)
fidelity = benchmark.get_fidelity_space().sample(random_state=0)
measurement = benchmark.evaluate(config=config, fidelity=fidelity)

print(measurement["perf"])
print(measurement["cost"])
print(measurement["hardware"])

Included Files

• croissant.json: MLCommons Croissant metadata for the dataset.
• metadata/croissant_records.csv: sample-level manifest with system, category, fidelity, CSV, log, and hardware references.
• metadata/croissant_records.parquet: Parquet version of the sample-level manifest.
• format.md: normalized CSV schema and artifact naming rules.
• experiment-data/README.md: full data package structure and extraction guide.
• data/: HF-native Parquet tables for benchmark rows, configs, metrics, and artifact references.
• artifacts/: indexed .tar.zst shards containing full log and hardware artifacts.
• example-data/: a small example subset for quick local inspection.

License

The dataset metadata and benchmark data are released under CC-BY-4.0. The code in the GitHub repository is licensed separately under GPLv3.