rohitMukhi/prompt_complexity_dataset

Prompt Complexity Dataset

A curated dataset designed to train and fine-tune lightweight models (such as DeBERTa-v3-base) for real-time Prompt Complexity Classification.

This dataset helps engineering teams implement intelligent LLM routing—dispatching simple queries to low-cost local models and routing high-complexity prompts to powerful frontier models, minimizing API costs and latency.

---

Dataset Structure

The dataset is stored in JSONL (jsonlines) format, where each line represents a single training record with a text prompt and its corresponding complexity tier label.

Data Fields

• text (string): The user prompt.
• label (integer): The complexity tier (0, 1, or 2).

Complexity Tiers

Label	Tier Name	Description & Examples
0	Low Complexity	Factual lookups, basic definitions, simple text summaries, short Q&A.
1	Medium Complexity	Writing standalone functions, light code refactoring, standard API integration, multi-step instructions.
2	High Complexity	System architecture design, debugging concurrency/race conditions, ambiguous bug reports.

---

Quick Start (Loading the Dataset)

You can load this dataset directly into your Python training pipeline using the Hugging Face datasets library:

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("your-username/prompt-complexity-dataset")

print(dataset["train"][0])
# Output: {'text': 'What is the capital of France?', 'label': 0}