regolo/brick-complexity-extractor

---
license: cc-by-nc-4.0
task_categories:
- text-classification
language:
- en
tags:
- complexity-classification
- llm-routing
- query-difficulty
size_categories:
- 10K<n<100K
---

# Brick Complexity Extractor Dataset

Training dataset for a query complexity classifier (easy/medium/hard) used in the Brick LLM routing system.

## Overview

- **76,831 samples** labeled by Qwen3.5-122B as LLM judge
- **3 classes**: easy (27.5%), medium (43.2%), hard (13.5%)
- **Splits**: train (65,307), val (7,683), test (3,841)

## Label Criteria

| Label | Description |
|-------|-------------|
| easy | 1-2 reasoning steps, surface knowledge, factual recall |
| medium | 3-5 reasoning steps, domain familiarity, multi-step derivation |
| hard | 6+ steps, deep expertise, graduate-level specialization |

## Format

Each line is a JSON object:
```json
{"text": "What is the capital of France?", "label": "easy"}
```