---
pretty_name: MuxaBenchGold
language:
- en
license: mit
size_categories:
- n<1K
task_categories:
- text-generation
tags:
- benchmark
- agentic-coding
- code-generation
- zig
- physics
- biomechanics
- systems
- security
- code
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train.jsonl
---

# MuxaBenchGold

Agentic-coding benchmark in pure **Zig**: 55 tasks from easy to expert.

- 26 base tasks: rendering, physics, ML kernels, systems, crypto
- 29 hardest expand tasks: physics + biological mechanics + Zig applications hitting AI weak spots (delay instability, friction flips, fluid-structure coupling, fatigue/history, topology change + determinism, energy drift, zero-alloc)
- Each task: human-style English `prompt` + ONE `expected_result` + simple score `0-10`, `pass_score 7`
- Every task requires: build with your local `zig version` (no fixed version), report `zig_version`, deterministic rerun hashes, zero allocations after init proven by counters

## Load

```python
from datasets import load_dataset
ds = load_dataset("muxapro/MuxaBenchGold")
print(ds["train"][0].keys())
# id, title, category, difficulty, prompt, expected_result, checks, score_max, pass_score, scoring
```

## Score

- Each `checks` entry is yes/no with numbers
- Sum to `score_max=10`, pass if `>=pass_score=7`
- Human + LLM-judge friendly: one result, one score

## Files

- `data/train.jsonl` — 55 lines, one task per line (runner)
- `MuxaBench-Gold.json` — pretty 922-line version for reading in editor
- Source of truth: local `Desktop/Muxa/MuxaBench-Gold.json`

## License

MIT