Minimum Risk Training (MRT) directly optimizes sequence‑level evaluation metrics, overcoming the limitations of token‑level maximum likelihood (Shen et al., 2016). Recent studies (Yang et al., 2024; Jinnai et al., 2025) have revived interest in risk‑based optimization for modern language models. In this work we apply MRT to power‑outage report generation for the Outage Data Initiative Nationwide (ODIN). Heterogeneous source reports are transformed into standardized XML compliant with CIM IEC 61968‑3. After integrating MRT with Qwen2.5‑7B‑Instruct, overall accuracy rises from 16.20% to 68.95%, demonstrating the power of sequence‑level optimization for domain‑specific structured generation. The training pipeline consists of: ① pre‑training with standard cross‑entropy; ② computing risk using sequence metrics such as BLEU and ROUGE; ③ minimizing expected risk via gradient descent. Experiments show MRT markedly improves correct XML tag generation and coverage of critical business fields.
Review: MRT proves highly effective for structured generation in niche domains, suggesting broader adoption for small‑model scenarios.