Developer Fine-Tunes 7B LLM for Khattak Pashto Dialect Using LoRA and Unsloth

A developer fine-tuned a Qwen2.5-7B large language model to adapt it toward the Khattak Pashto regional dialect, using a technique called LoRA (Low-Rank Adaptation) alongside the Unsloth training framework. LoRA works by keeping the original model weights frozen and training only a small set of additional parameters, making the process far less resource-intensive than full fine-tuning. Unsloth complemented this by reducing memory usage and speeding up training, making the experiment feasible without access to large GPU clusters. Over the course of training, the model's loss dropped significantly from 3.44 to 0.22, though the developer noted that real-world evaluation remains essential beyond training metrics alone. The project, documented as Khatta-ka-LLM, highlights how parameter-efficient fine-tuning techniques are lowering the barrier to LLM experimentation for developers with limited infrastructure.