Fine-tuned Qwen on my 16gb GPU to respond without the letter e
Summary
The article documents fine-tuning Qwen-1.7B locally to output lipograms (text without the letter 'e'), comparing masked and RL-based approaches, and detailing memory-optimization strategies (LoRA, gradient checkpointing, fused cross-entropy). It presents experimental results showing improvements in usable responses after RL, along with practical observations and takeaways for reproducing local LLM fine-tuning on commodity hardware.