125m legal SLM with RAFT (retrieval-augmented fine-tuning, Zhang et al. 2024): base -> QA SFT -> instruction SFT -> DPO v2 -> RAFT. Trained on ~14k examples of golden passage + 3 distractors (golden position uniform, 20% no-golden refusals): answer from the right passage, ignore the noise, refuse when the context lacks the answer.