Fine-tuned Qwen on my 16gb GPU to respond without the letter e (kevin-wu.ca)

A developer successfully fine-tuned the Qwen3-1.7B language model on a personal 16GB GPU to generate text that omits the letter 'e', a challenging task given how language models process text in tokens rather than by specific characters. Inspired by the concept of lipograms, the developer aimed to create coherent responses to everyday prompts without using 'e', demonstrating a novel approach to both training and evaluation. Initial attempts reflected the model's struggle with the task; thus, a combination of supervised fine-tuning (SFT) and reinforcement learning (RL) was employed to improve model performance. This project is significant for the AI/ML community as it explores unique constraints in model outputs, enriching the understanding of language generation capabilities. Early results showed usable output rates of just 26%, but through refined training practices, including using a custom dataset of ‘e’-free answers, the model achieved up to 60.5% usability after RL training. This success highlights the potential for innovative training methodologies and the adaptability of language models, pushing towards more creative and constrained forms of AI text generation.