LoRA Fine-Tuning
What I built
Data curation/decontamination, completion-only masking, real LoRA training, adapter merge, GGUF export, local serving and evaluation.
Result
Produced adapter, merged-model and GGUF artifacts from a partial Qwen 2.5 0.5B CPU training run, connecting data preparation to local serving.
Evaluation setup and selected results
A Qwen 2.5 0.5B CPU LoRA run used 244 examples and 19 optimizer steps (0.32 epoch), producing adapter, merged-model and GGUF artifacts. The workflow connects data curation and completion-only training to export, local serving and evaluation. It fine-tunes an existing open-source model.
Compare the evaluation evidence →
About the evidence
Based on source review and saved project artifacts. Recorded measurements apply to the stated model, dataset and configuration; they are not claims of a current production deployment.