SCPe_Dataset_2015-2020

A structured dataset of forensic reports on violent deaths produced by the Crimes against People Team (Seção de Crimes Contra a Pessoa - SCPe), from the Forensic Institute of the Civil Police of Distrito Federal (PCDF), Brazil. Overview This repository contains data and scripts used to build a structured forensic dataset from originally unstructured textual reports. Coverage period: 2015–2020. Reports collected: 4,396. Reports successfully processed: 4,362. Case scope: homicide, suicide, accident, and undetermined manner of death. The goal is to support statistical analysis and NLP applications in a real forensic context while preserving the variability and uncertainty that are typical in violent-death investigations. Language note: the project is international, but the dataset itself is in Brazilian Portuguese (pt-BR). Repository structure SCPe-Dataset/ ├── SCPe_Dataset.csv.zip # compressed final structured dataset ├── loader.py # utility to load the CSV with pandas └── README.md Dataset schema The file SCPe_Dataset.csv contains summarized text fields and categorical variables. Text fields (semantic summary) texto_original: concise factual summary of the report. descricao_local: description of the event location. vestigios_local: relevant material traces at the scene. vestigios_cadaver: traces and observable conditions on the body. exames_complementares: additional forensic/laboratory examinations (when available). Main categorical fields maneira_morte: homicídio, suicídio, acidente, indeterminada (and real-world variations present in the data). instrumento: arma de fogo, perfurocortante, contundente, constrição cervical, etc. local_evento: whether the body was found at the event location. cadaver_local: presence of a body at the examined location (e.g., SIM/NÃO). sangue_local: presence of blood at the location. tipo_local: indoor/outdoor environment (or undetermined). preservacao: scene preservation status before forensic examination. agente_estatal: involvement of a state agent (sim/não/indeterminado). How the dataset was built (summary) Report collection from institutional archives. PDF-to-text conversion using pdfplumber. Text preprocessing (normalization of line breaks and spacing). LLM-based structured extraction via prompting, without external domain knowledge. Post-processing and consolidation into tabular (CSV) format. Quick start 1) Convert PDFs to TXT python conversion_txt.py 2) Generate the structured dataset from TXT files python script_dataset.py 3) Load into a pandas DataFrame from loader import load_dataset df = load_dataset() print(df.shape) Loader (loader.py) First, to use the loader.py, the SCPe_Dataset.csv.zip should be uncompressed in the same local. The loader.py file centralizes dataset loading with simple path validation. Main function load_dataset(path=None, **read_csv_kwargs) -> pandas.DataFrame If path is not provided, it defaults to dataset/dataset.csv. If the file does not exist, it raises FileNotFoundError. Any extra arguments are forwarded to pandas.read_csv (for example, sep, encoding, usecols). Usage examples Default loading from loader import load_dataset df = load_dataset() Load from a specific path from loader import load_dataset df = load_dataset("dataset/dataset.csv") Pass read_csv parameters from loader import load_dataset df = load_dataset(usecols=["maneira_morte", "instrumento"], encoding="utf-8") Run the loader directly from the terminal python loader.py This command loads the default dataset and prints: number of rows and columns; first rows (head) for quick inspection. Requirements Python 3.10+ Main libraries: pandas pdfplumber openai Suggested installation: pip install pandas pdfplumber openai Methodological and ethical considerations The dataset is intended for research and computational method development. The material represents real forensic scenarios and may contain uncertainty inherent to investigative processes. Users must comply with applicable legal, ethical, and institutional requirements for forensic data usage. The dataset should not be used in isolation for forensic, police, or judicial decision-making. All personal names were removed through an automated anonymization process prior to dataset release, ensuring compliance with data protection requirements. Potential applications Descriptive statistics for violent death patterns. Predictive modeling and classification of forensic events. Information extraction and summarization in forensic documents. NLP robustness studies on noisy and ambiguous data. Citation (provisional) If this repository is useful to your research, please cite the article associated with the dataset: Carneiro, L. D.; Vidal, F. B. A structured dataset of forensic reports from violent death investigations in Brazil (2015–2020). (project manuscript) Contact Leandro Dias Carneiro — [e-mail] Flavio de Barros Vidal — [e-mail]