qwen2.5-7b-advocate-sft-lora-ep1

LoRA adapter for Qwen2.5-7B-Instruct, trained with QLoRA supervised fine-tuning (epoch 1) on question-answer data about Kenya's Advocates Remuneration Order and related court rulings. Training data was written by a teacher model (gpt-oss-20b).

This is the "SFT" system in a pre-registered evaluation of pipeline engineering versus fine-tuning. On 69 unseen-ruling questions it scored 2.9 points below the base model (95% interval -10.1 to +4.9), so no benefit is detectable. See the project repository for the study and code: https://github.com/mikekeror-rgb/advocate_renumeration_assistant_v1

Research artefact, not legal advice. The DPO adapter built on this one is at mikekeror/qwen2.5-7b-advocate-dpo-lora.