touching-machines/prompt-optimization-pii-masking-2k

🚀 Hugging Face Dataset Plan: PROMPT_OPTIMIZATION & PII_MASKING

This dataset focuses on two core areas critical for modern enterprise applications: Prompt Engineering (Token/Jargon Efficiency) and Sensitive Data Management (PII Masking).

1\. Dataset Title (Repository Name)

Hugging Face titles should be short, descriptive, and task-oriented. Using an English, jargon-rich name is recommended to target the international NLP/LLM community.

2\. Dataset Description

This section is vital for explaining the dataset's value and potential use cases.

A. Short Description (Abstract)

"A comprehensive, multilingual (English), and synthetically generated instruction dataset containing 2015 unique examples focused on two critical tasks for Large Language Models (LLMs) in Enterprise/SaaS environments: 1) Prompt Optimization (maximizing token and jargon efficiency) and 2) PII Masking (applying secure, context-aware redaction across international formats and complex technical scenarios)."

B. Purpose and Use Cases

This dataset is uniquely designed to serve the following key NLP/LLM applications:

1. 1.Instruction Tuning / Fine-Tuning:
2. 2.To enhance LLMs' ability to compress long and verbose prompts (original_input) into technically efficient, jargon-rich, and concise instructions (target_output).
3. 3.To help LLMs understand and correctly use technical, FinOps, DevOps, and Project Management acronyms (e.g., RoI, MTTD, DORA, BCP, GTM, P95 Latency).
4. 4.Information Extraction & Security:
5. 5.To train LLMs to recognize PII in complex, nested, or international formats (e.g., Spain NIE, Netherlands Sofi, various bank SWIFT codes, API keys, IP addresses, different phone number lengths).
6. 6.To improve the LLM's capability to accurately detect and redact sensitive information using predefined masking tags (e.g., `[CUSTOMER_NAME]`, `[ID_NUMBER]`, `[API_KEY]`).
7. 7.Benchmarking:
8. 8.Can be used as a challenging benchmark set to measure and compare the performance of various LLMs on high-stakes tasks related to Prompt/Token efficiency and PII Security/Masking accuracy.

C. Dataset Structure (Features)

You must clearly define what each column (feature) represents:

3\. Usage Example (Code Snippet)

Include a simple Python code example to show users how to load and interact with the dataset:

from datasets import load_dataset

# Load the dataset (Replace YOUR_USERNAME with your actual Hugging Face username)
dataset = load_dataset("YOUR_USERNAME/prompt-optimization-pii-masking-2k")

# Print a Prompt Optimization example
print("--- PROMPT OPTIMIZATION EXAMPLE ---")
opt_example = dataset['train'].filter(lambda x: x['task_type'] == 'PROMPT_OPTIMIZATION')[0]
print(f"Original Input (Verbose Prompt): {opt_example['original_input']}")
print(f"Target Output (Optimized): {opt_example['target_output']}")

# Print a PII Masking example
print("\n--- PII MASKING EXAMPLE ---")
pii_example = dataset['train'].filter(lambda x: x['task_type'] == 'PII_MASKING')[0]
print(f"Original Input (Contains Sensitive Data): {pii_example['original_input']}")
print(f"Target Output (Masked Data): {pii_example['target_output']}")

license: mit