Fine-Tuning Models for Latin American Firearm Safety

In Latin American countries, where trust in law enforcement is at an all-time low and corruption runs rampant, the perceived necessity for guns is dangerously high. Unfortunately, while the media is an invaluable information outlet, its constant over-exemplification of violence and glorification of killers leads too many down the wrong path. Accurately identifying the amount of gun violence promoted in the media can lead to adjustments in media portrayal of shootings, creating less “idolization” of mass killers.
To efficiently measure the frequency of gun violence appearing in the media, an efficient categorization system is needed. This paper uses the large language model LLaMA-2’s ability to classify news reports based on the mention of firearm-related concepts within them. A pretrained LLaMA-2 with 7B parameters model was fine-tuned via a web-scraped, balanced dataset containing 531 training samples and 133 test samples. After being cleaned from mistakes, the dataset was organized based on whether the input values used words—in either English, Spanish, or Portuguese—related to firearms, labeled through human efforts with “True” indicating the mention of guns, or “False” indicating the absence of guns in the text. The model was trained to achieve specialization in news classification, successfully classifying firearm mentions. The results of this experiment can be further used in research regarding the best ways to control firearm popularity, as news and media articles contribute a large part to firearm sales, especially when they are glorified as the perfect self-defense weapon.

Built With

• python