In today's digital age, misinformation spreads like wildfire across social media and the internet, making it harder than ever to distinguish fact from fiction. This poses a huge challenge for fact-checkers who need to identify which claims are actually worth investigating. This is where AI could step in. Our team, FactFinders, explored how open-source Large Language Models (LLMs) can be used to automatically detect check-worthy claims from political speeches. We dove into a dataset of over 24,000 sentences, many of them quite short and not very informative, and found that the data itself posed a challenge. So we developed a two-step "data pruning" process to filter out the noise and focus on the most informative parts. This involved identifying sentences with named entities, informative verbs, or sufficient length, then using a technique called Condensed Nearest Neighbour to balance the remaining data. We tested eight different LLMs, including Llama2, Llama3, Mistral, Mixtral, Phi-3, Falcon, and Gemma. We fine-tuned each of these models on the data using carefully crafted prompts to guide their analysis. Interestingly, we found that our fine-tuned Llama2-7b model, using our data pruning method, showed the most promise in detecting check-worthy claims, even though it trained on a substantially smaller dataset than the original. In fact, it only used 44% of the original data, which greatly reduced the time and resources needed for training. This not only suggests that bigger isn't always better when it comes to training these models, but also emphasizes the importance of data quality. Our FactFinders submission using the fine-tuned Llama2-7b model achieved the top spot in the CheckThat! 2024 competition, showing that open-source LLMs can provide accurate results in check-worthy claim detection. While the LLMs showed great promise, we encountered a few challenges. Some models required significant computing resources to fine-tune, and there was inconsistency in predictions—sometimes generating different results for the same input. This underscores the need for further research into efficient training methods and reliable prediction techniques. Our work is a significant step towards automating fact-checking, helping separate truth from falsehood in the ever-expanding ocean of information online. As AI models evolve, they hold even greater potential for this vital task, helping us navigate the world of information more effectively than ever before.