Joe-Zeloof/DS_Final_Project_JZ_RF_Parking_Ticket_AI_DataSet

Synthetic Parking Ticket Cases (10,000)

This repository contains a synthetic dataset of parking ticket “case narratives” generated with the Qwen/Qwen3-0.6B language model for a data science course project.
Each record describes a parking offense in exactly ~5 short lines and includes contextual details such as location, time, car involved, how the offense was recorded, the driver’s explanation, and signage clarity.

Important: This dataset is fully synthetic and is not based on real legal records. It should not be used for legal advice or real-world enforcement decisions.

---

Dataset Contents

Files

• parking_cases_final.csv — final dataset after EDA, enriched with derived columns
• eda......ipynb — exploratory data analysis notebook
• synthetic......ipynb — synthetic data generation notebook

Columns (enriched CSV)

• case_id (int): Unique identifier (1..10,000)
• offense_description (string): A short narrative (~5 lines) describing the parking offense
• Text stats
  • word_count, char_count, line_count
  • is_exact_5_lines (boolean)
  • has_non_ascii (boolean)
• Heuristic extractions (from text)
  • recording_method (ALPR / CCTV / Photos / Handheld/Officer / Unknown)
  • ` car_make_guess

Note: The extracted fields are heuristics (keyword/regex-based), and will contain some noise. They are intended to enrich EDA and enable structured modeling.

---

How the Data Was Generated

• Model: Qwen/Qwen3-0.6B
• Generation strategy:
  • Prompted to produce exactly 5 short lines per case
  • Batched GPU generation for speed (e.g., multiple prompts per model.generate call)
  • Saved in batches to reduce failure risk and allow resuming
• Cleaning:
  • Filtered out low-quality outputs such as strings dominated by punctuation (e.g., "??????") using an alphabetic character ratio threshold.
  • (Optional) Filtered out overly short narratives (e.g., < 20 words) for dataset consistency.

---

EDA Summary

The included notebook eda.ipynb performs exploratory data analysis to assess:

• Dataset size, missing values, duplicates
• Text length distribution:
  • characters per case
  • words per case
  • line count per case
• Data quality checks:
  • % of cases with exactly 5 lines
  • % of cases containing non-ASCII characters
  • detection of low-information “junk” outputs (punctuation-heavy strings)
• Heuristic extraction summaries:
  • recording method frequency (CCTV/photos/officer device/ALPR)
  • approximate car make distribution (keyword-based)

Key Quality Notes

• Synthetic text models occasionally produce corrupted outputs (e.g., punctuation-only strings).
  To address this, a cleaning step keeps cases whose alpha_ratio exceeds a chosen threshold (e.g., >= 0.50 in our case - meaning a majority of non alphabet characters).
• Extracted columns in the enriched dataset are approximate and should be treated as noisy features.

---

Example Record (illustrative)

A typical offense_description includes:

• A specific city + street/landmark and the exact parking context (e.g., hydrant, loading zone)
• A realistic date/time
• Vehicle details (year/make/model/color + plate-like string)
• Evidence method (officer handheld device, photos, CCTV, ALPR)
• Driver’s explanation (delivery, emergency, couldn’t find parking, etc.)
• Whether signage/road markings were clearly visible

---

Intended Use

• EDA practice (text analytics, cleaning, and quality checks)
• NLP feature engineering (length features, keyword extraction, weak labels)
• Baseline modeling tasks (e.g., classify violation type or signage clarity from text)
• Demonstrating a full synthetic data pipeline: generation → cleaning → EDA → publishing

---

Limitations & Ethics

• Not real data: All cases are synthetic and may not reflect real enforcement language.
• Bias & realism: Content depends on prompt design and model behavior. The dataset may contain stereotypes or unrealistic details.
• No legal validity: Do not use for legal decisions, policy, or enforcement.

---

Loading the Data

With pandas

import pandas as pd
df = pd.read_csv("parking_cases_final.csv")