Worked on stabilizing model training and improving codebase consistency in the liguodongiot/transformers repository, focusing on deep learning and machine learning workflows. Addressed a bug in the Qwen2_5_VLForConditionalGeneration model by correcting vocab_size handling within the loss computation, ensuring the correct configuration value was used. This targeted fix reduced training noise and improved the accuracy of loss metrics, directly impacting the reliability of evaluation results. Propagated similar adjustments to related models, promoting uniformity in loss calculations across the codebase. Utilized Python for model training and bug resolution, contributing to more stable pipelines and maintainable deep learning infrastructure.