Self-hosted alternatives to lora-qlora-recipes

L

lora-qlora-recipes

Configure LoRA and QLoRA supervised fine-tuning

lora-qlora-recipes is an agent skill from the wshobson/agents collection for configuring LoRA and QLoRA supervised fine-tuning with current best-practice settings. LoRA trains a small set of adapter weights instead of the whole model, and QLoRA does the same on a quantized base model to save memory. The skill covers the decisions that shape a LoRA run: which modules of the model to target, how to size the rank and alpha of the adapter, what learning rate to use, and when QLoRA gives real headroom on limited hardware compared with when it only adds risk. It is used when writing or reviewing a LoRA or QLoRA training configuration. The skill assumes the routing step has already picked supervised fine-tuning, because the data is demonstrations rather than preference pairs or a verifiable reward signal. Preparing the dataset and checking its quality are a separate concern, covered by the dataset-curation skill. The guidance helps an agent produce a configuration a team can review and then run with its own training tools. As with any fine-tuning, results should be measured against an evaluation baseline before the adapter is used. The skill is installed with the skills CLI from the wshobson/agents repository. Its skills.sh page showed about 2.4K installs and passing results from three security audits at the time of the scan.