ai-nikolai/ScienceKernelBench

Science KernelBench

A science-focused task set for robust-kernelbench. Each task provides a PyTorch reference implementation of a numerical primitive; the candidate must produce a faster, numerically equivalent GPU kernel.

This directory is the source of truth. to_hf.py expands the manifest into a Hugging Face dataset; the generated export is not committed.

Current starter set: 22 tasks across physics, chemistry, biology, statistics, mathematics and economics.

level	focus	tasks
level_1	single stencil, reduction or short elementwise expression	8
level_2	all-pairs interactions, batched linear algebra, multi-state updates	7
level_3	sequential dynamic programs, FFTs, tree backward induction	7

Each task carries domain, subdomain, level, tags and a source block. The reference files contain a docstring with the domain, the problem solved and the canonical paper/source.

Task list

id	task	domain	level
101	pairwise_rbf	statistics	1
102	laplacian_2d	physics	1
103	heat_step_2d	physics	1
104	lennard_jones_energy	chemistry	1
105	lotka_volterra_step	biology	1
106	gaussian_kde	statistics	1
107	horner_eval	mathematics	1
108	black_scholes_call	economics	1
201	nbody_acceleration	physics	2
202	coulomb_energy	chemistry	2
203	hodgkin_huxley_step	biology	2
204	kalman_update	statistics	2
205	graph_laplacian	mathematics	2
206	portfolio_variance	economics	2
207	wave_step_2d	physics	2
301	needleman_wunsch_score	biology	3
302	fft_convolution	mathematics	3
303	binomial_option_crr	economics	3
304	split_step_schrodinger	physics	3
305	ewald_real_space	chemistry	3
306	gaussian_process_posterior	statistics	3
307	floyd_warshall	mathematics	3

Versioning

Current release: 0.1.0 (schema_version: kernelbench-v1), compatible with robust-kernelbench >= 0.2.0.

The dataset version is independent of the package version:

• PATCH — documentation, metadata or provenance fixes with no scoring change.
• MINOR — new tasks, domains or levels; existing tasks unchanged.
• MAJOR — removal/renumbering, reference-output changes, or benchmark-split changes.

Every task has a permanent problem_id, a stable task_key, a task_version and a reference_hash. Removed tasks are marked retired; IDs are never reused. Pin production runs to an HF tag or commit:

robust-kernelbench \
  --dataset ai-nikolai/ScienceKernelBench \
  --dataset_revision v0.1.0 \
  --dataset_split level_1 \
  ...

Schema

Required columns:

column	type	meaning
problem_id	int	unique task id
name	string	filesystem-safe task name
code	string	self-contained reference source

The code must define:

class Model(nn.Module): ...
def get_inputs(): ...
def get_init_inputs(): ...

Optional columns understood by the harness:

column	meaning
level	dataset split / difficulty grouping (1-3 here)
domain	physics, chemistry, biology, statistics, mathematics, economics
subdomain	finer label, e.g. molecular dynamics, PDE, bioinformatics
filter	include in the representative subset
task_type	pytorch_reference or kernel_reference
backend	cuda, triton, ...
precision	fp32, fp16, bf16
tags	searchable labels
source	provenance / license metadata

Local layout

data/science_kernels/
├── README.md                  # this dataset card
├── manifest.json              # versioned task list and provenance
├── CHANGELOG.md               # dataset release history
├── requirements.txt           # minimal HF/export dependencies
├── kernels/
│   ├── level_1/               # 8 tasks
│   ├── level_2/               # 7 tasks
│   └── level_3/               # 7 tasks
└── to_hf.py                   # per-level export / optional push_to_hub

Minimal dependencies

pip install -r requirements.txt

torch is optional and only needed for --execute smoke tests; it is listed as a comment in the requirements file so the dataset can be loaded and exported on a lightweight machine.

Validate locally

# Schema only
robust-kernelbench-dataset validate \
  --dataset_path data/science_kernels \
  --dataset_format directory

# Counts by level and domain, and a browsable list
robust-kernelbench-dataset stats \
  --dataset_path data/science_kernels \
  --dataset_format directory
robust-kernelbench-dataset list \
  --dataset_path data/science_kernels \
  --dataset_format directory --domain physics

# Also run one forward pass per task
robust-kernelbench-dataset validate \
  --dataset_path data/science_kernels \
  --dataset_format directory \
  --execute --device cpu

Use with robust-kernelbench

Evaluation only:

robust-kernelbench-eval \
  --dataset_path data/science_kernels \
  --dataset_format directory \
  --dataset_split level_1 \
  --experiment science_exp --trial 1

Agentic generate -> compile -> evaluate:

robust-kernelbench \
  --dataset_path data/science_kernels \
  --dataset_format directory \
  --dataset_split level_1 \
  --dataset_name science-kernels \
  --version science_v0 \
  --model_name Qwen/Qwen3-Coder-30B-A3B-Instruct \
  "1 1 kernelbench"

Export for Hugging Face

python data/science_kernels/to_hf.py --output /tmp/science_kernels_hf

# huggingface_hub >= 1 uses the `hf` command:
hf upload ai-nikolai/ScienceKernelBench /tmp/science_kernels_hf --repo-type dataset

# Works with any huggingface_hub version:
python data/science_kernels/upload_to_hf.py \
  --path /tmp/science_kernels_hf

Or export and push in one step:

python data/science_kernels/to_hf.py \
  --output /tmp/science_kernels_hf \
  --push

Provenance and licensing

Only add kernels you have the right to redistribute. Record the source repository, commit, and license in each task's source block. Original tasks written for this dataset are released under AGPL-3.0, matching the repository. If you combine sources with different licenses, use the strictest compatible license for the combined dataset and document it in this card.

Citation

If you use this dataset, please cite robust-kernelbench:

@misc{rozanov2026robustkernelbench,
  title        = {Robust KernelBench: A Robust Harness for GPU Kernel Generation and Evaluation},
  author       = {Rozanov, Nikolai},
  year         = {2026},
  howpublished = {\url{https://github.com/ai-nikolai/robust_kernelbench}},
  note         = {ScienceKernelBench dataset, version 0.1.0}
}

The dataset card and version should also be cited when reporting benchmark results, together with the exact dataset_revision used.