metadata
license: mit
language:
- en
task_categories:
- text-generation
tags:
- benchmark
- code
- evaluation
- debugging
size_categories:
- n<1K
Code Stress Benchmark
50-task evaluation benchmark for code-focused LLMs. Built to test whether a fine-tune holds up across realistic workflow scenarios — not just isolated bug-fix puzzles.
Categories
| Category | Tasks | What it tests |
|---|---|---|
| easy_bug | 5 | Common Python/JS pitfalls |
| tricky_bug | 5 | Race conditions, default args, scoping |
| medium_script | 10 | Multi-step Python/JS/SQL/Bash |
| hard_algorithm | 10 | LRU cache, DP, graph traversal, LCA |
| edge_cases | 5 | Boundary conditions, error handling |
| architecture_review | 5 | Code review on small services |
| engineering_judgment | 5 | "Should I do X or Y?" prompts |
| design_reasoning | 5 | Open-ended system design |
Format
{
"id": "bug_easy_01",
"category": "easy_bug",
"language": "python",
"prompt": "Find and fix the bug:\n\n```python\ndef average(nums):\n return sum(nums) / len(nums)\n\nprint(average([]))\n```"
}
Use
Designed for A/B comparison: run the same task through base and tuned models, then measure:
- length ratio (lora / base) — does the tune produce more focused answers?
- lazy regression — does the tune drop code blocks the base produced?
- correctness — manual review
Used to evaluate NecroMOnk/Residual and NecroMOnk/Tersa.
from datasets import load_dataset
ds = load_dataset("NecroMOnk/code-stress-bench")