# Vision-language instruction fine-tuning

4-bit quantized version based on Qwen2-VL-7B model, trained using Unsloth and Huggingface TRL library, achieving 2x inference speedup