You need to agree to share your Hugging Face information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
Access requires explicit approval from the owner. By requesting access you agree NOT to use this dataset, in whole or in part, for training, fine-tuning, evaluating, or benchmarking any AI/ML model, and to make no other use without prior written permission from the owner.
Log in or Sign Up to review the conditions and access this dataset content.
v4
A benchmark of 244 low-level kernel tasks for code generation, split into 189 train / 55 holdout. Each task is a self-contained programming problem: a function contract, a natural-language prompt, a scalar reference implementation, a test harness, and near-miss variants used to catch almost-correct solutions.
Splits
| Split | Tasks |
|---|---|
| train | 189 |
| holdout | 55 |
Row fields
| Field | Description |
|---|---|
task_id |
Unique task name |
level |
Difficulty level (L1/L2/L3) |
domain |
Task domain |
category |
Task category |
dtype |
Primary data type |
caps |
Required hardware capabilities |
signal |
One-line characterization (e.g. compute- vs bandwidth-bound) |
baseline_cycles |
Reference baseline cost |
tags |
Keyword tags |
prompt |
The task prompt (Markdown) |
kernel_api_h |
The function contract header |
baseline_c |
Scalar reference implementation |
nearmiss_files |
Near-miss variant filenames included in the task folder |
spec_json |
Full task spec as JSON |
The full per-task files (prompt, spec, baseline, harness, near-miss variants)
also live unpacked under tasks/<task_id>/.
Notes
- Source files are original. This dataset does not include any vendor SDK, toolchain, headers, or binaries — those must be obtained separately.
License
All rights reserved. Any use requires prior written permission from the
owner — see LICENSE. To request permission, contact the owner via
the Hugging Face profile hitmanonholiday.
- Downloads last month
- 11