Datasets:

sinatras
/

pmpp-eval

pretty_name: pmpp-eval
license: mit
task_categories:
  - text-generation
  - question-answering
language:
  - en
tags:
  - pmpp
  - cuda
  - coding
  - mcq
  - qa
size_categories:
  - n<1K
configs:
  - config_name: qa
    data_files: pmpp_qa.jsonl
    default: true
  - config_name: coding
    data_files: pmpp_coding.jsonl

PMPP Dataset

This repository provides two CUDA-focused datasets prepared by Sinatras and sponsored by Prime Intellect. Both datasets are based on Programming Massively Parallel Processors (4th Ed.) with additional coding evaluation harnesses at https://github.com/SinatrasC/pmpp-eval to be used by PMMP env in prime-environments.

Overview

Languages: English
License: MIT
Curated by: Sinatras (https://github.com/SinatrasC)
Sponsored by: Prime Intellect
Derived from: PMPP 4th Edition (Kirk & Hwu)

Dataset Details

pmpp_qa

Composition: 61 MCQ + 77 short-answer items.
Fields: chapter, exercise, type, question, answer, explanation, topic_tags, and optional choices.
Topics emphasize CUDA indexing, occupancy, memory hierarchy, MPI, and dynamic parallelism.

pmpp_coding

53 coding tasks. Every entry corresponds to evaluation-tasks/<id>/student_kernel.cu and includes runner metadata.
Fields: id, task_dir, student_file (always student_kernel.cu), optional test targets/executables, and the trimmed CUDA skeleton.
Some tasks export host wrappers (e.g., device property collection, one-pass radix) rather than __global__ kernels. Tests under the same directory call the exported symbols.

Coding Dataset Evaluation Sample

Dataset was evaluated using the coding evaluation harness within the PMPP env on prime-environments (https://github.com/PrimeIntellect-ai/prime-environments/)

Model Performance

Model	Total Tasks	Success Rate	Rollouts
Qwen/Qwen3-Next-80B-A3B-Thinking	53	24.5% (39/159)	3 per task

Top Performing Tasks

Task	Success	Description
ch02-vecadd-single-turn	3/3	Vector addition kernel
ch03-rgb2gray-single-turn	3/3	RGB to grayscale conversion
ch09-histogram-naive-single-turn	3/3	Histogram computation
ch09-histogram-shared-single-turn	3/3	Histogram computation
ch14-spmv-csr-thread-per-row-single	3/3	Sparse matrix-vector multiply
ch14-spmv-coo-single	3/3	Sparse matrix-vector multiply
ch14-spmv-ell-single	3/3	Sparse matrix-vector multiply
ch18-energy-gather-coarsened-single	3/3	Energy simulation kernel

Most Challenging Areas (0% Success)

Challenge Category	Failed Tasks	Examples
Matrix Operations	6 tasks	Matrix multiplication variants, tiled algorithms
Advanced Algorithms	8 tasks	Sorting, reduction, merge operations
Memory Optimization	12 tasks	Shared memory, coalescing, thread coarsening
MPI Integration	3 tasks	Multi-GPU communication patterns
Dynamic Parallelism	3 tasks	Parent-child kernel launches
Graph Algorithms	3 tasks	BFS, sparse data structures

Intended Use

pmpp_qa: Evaluate or fine-tune GPU aware assistants on conceptual CUDA/MPI reasoning.
pmpp_code: Evaluate or fine-tune code-generation capabilities.

Limitations

Specialized, CUDA-focused coding evaluation harness only.
Some coding tasks require runtime configuration (e.g., enabling device heap). The pmpp-eval harness handles those details.

Acknowledgements & Citation

Grateful acknowledgment to Prime Intellect for sponsoring this release and to the PMPP community for foundational materials. Additional inspiration and reference code were drawn from the open solution set at https://github.com/tugot17/pmpp. If you build on these datasets, please cite both sources:

@book{kirk2016programming,
  title     = {Programming Massively Parallel Processors: A Hands-on Approach},
  author    = {Kirk, David B. and Hwu, Wen-mei W.},
  edition   = {4th},
  year      = {2016},
  publisher = {Morgan Kaufmann}
}

@misc{pmpp_eval,
  author = {Sinatras},
  title  = {pmpp-eval},
  year   = {2025},
  url    = {https://github.com/SinatrasC/pmpp-eval}
}

For questions or contributions, open an issue in https://github.com/SinatrasC/pmpp-eval.