Datasets:

myhapro
/

MuxaBench

pretty_name: MuxaBench
language:
  - en
license: other
license_name: muxa2
license_link: LICENSE
size_categories:
  - n<1K
task_categories:
  - question-answering
  - other
tags:
  - benchmark
  - evaluation
  - code
  - systems-programming
  - zig
  - coding-benchmark
  - llm-benchmark
  - rubric
configs:
  - config_name: default
    data_files:
      - split: test
        path: MuxaBench.jsonl
dataset_info:
  features:
    - name: id_aa
      dtype: string
    - name: title
      dtype: string
    - name: category
      dtype: string
    - name: prompt
      dtype: string
    - name: system_prompt
      dtype: string
    - name: rubric
      dtype: string
    - name: expected_deliverables
      dtype: string
    - name: reference_files
      dtype: string
  splits:
    - name: test
      num_bytes: 270174
      num_examples: 26
  download_size: 270174
  dataset_size: 270174

MuxaBench: Systems Programming Benchmark

MuxaBench is an evaluation benchmark designed to assess LLMs and coding agents on 26 staff/principal-level systems programming tasks in pure Zig 0.15.1.

Benchmark Overview

The benchmark contains 26 tasks divided into 5 core engineering domains:

Rendering and content pipelines (6 tasks): ReSTIR DI on BVH, perceptual 8 bpp texture codec, analytic-AA vector rasteriser, 10 km² city + navmesh, HZB occlusion culling.
Physics and motion simulation (6 tasks): GJK/EPA CCD rigid body, XPBD rope/cloth, active ragdoll, full-body IK, 128³ Eulerian multigrid smoke solver, quadcopter simulator.
Numerical and ML kernels (5 tasks): Fused transformer step, Int8 GQA engine with paged KV-cache, runtime SPIR-V shader generator, FlashAttention causal attention, spatial audio synthesizer.
Systems infrastructure (5 tasks): Crash-safe WAL KV engine, deterministic matching engine (10M ops/s), x86_64 microkernel slice (4-level VMM/page allocator), Raft under 30% packet loss, JIT compiler with generational GC.
Security and cryptography (4 tasks): Constant-time ML-KEM-768, ML-DSA-65 signatures, R1CS SHA-256 prover, static PE heuristic analyzer.

Evaluation Methodology (Score Against Rubric)

Each task in MuxaBench is scored against a multi-criteria rubric with discrete scoring levels (0 to 2 or 3) across ~6 explicit criteria per task. Models are evaluated on:

Strict correctness and mathematical formulation
Enforcing zero dynamic allocation in hot loops after initialization
Deterministic hashing across runs and threads
metric_provenance: verifying that reported metrics trace directly to executable test harness code