Developed an end-to-end video-language model fine-tuning pipeline for the huggingface/trl repository, enabling researchers to download videos, process time-coded subtitles, and prepare datasets for Qwen2-VL training. The solution incorporated configurable training parameters, quantization, and LoRA to support efficient and reproducible experimentation with video-language models using Python and PyTorch. Additionally, improved the liguodongiot/transformers repository by enhancing the Idefics3 processor’s robustness, specifically addressing text-only input scenarios. This involved implementing explicit error handling and validation to prevent silent failures, thereby increasing reliability for downstream machine learning workflows. Work demonstrated depth in computer vision, data processing, and unit testing practices.