Enhancing Small Language Models for Power Outage Report Generation via Minimum Risk Training

Forum topic · 小凯 · 2026-09-25

Summary

This paper applies Minimum Risk Training (MRT) to power outage report generation, enabling small language models to directly optimize sequence-level evaluation metrics instead of relying solely on token-level maximum-likelihood objectives. The task targets the Outage Data Initiative Nationwide (ODIN), transforming heterogeneous outage reports into standardized XML compliant with CIM IEC 61968-3. While MRT was introduced a decade ago for neural machine translation, recent research has shown renewed potential for risk-based optimization in modern language models. The authors—Hung Phan, Waqwoya Abebe, Youssef Hussein, Supriya Chinthavali, Dalton Lunga, and Ali Jannesari—demonstrate that fine-tuning Qwen2.5-7B-Instruct with MRT raises overall accuracy from 16.20% to 68.95%, a substantial improvement that highlights the effectiveness of sequence-level optimization for domain-specific structured generation tasks. The work is available on arXiv (2609.27197) and offers a practical recipe for using compact language models in utility-domain document standardization.

Paper Overview

Field: Machine Learning Authors: Hung Phan, Waqwoya Abebe, Youssef Hussein, Supriya Chinthavali, Dalton Lunga, Ali Jannesari arXiv: 2609.27197

Abstract

Minimum Risk Training (MRT) enables neural machine translation models to directly optimize sequence-level evaluation metrics instead of relying only on token-level maximum-likelihood objectives (Shen et al., 2016). Although introduced a decade ago, recent work shows renewed potential for risk-based optimization in modern language models (Yang et al., 2024; Jinnai et al., 2025).

The authors apply MRT to power outage report generation for the Outage Data Initiative Nationwide (ODIN), transforming heterogeneous reports into standardized XML compliant with CIM IEC 61968-3.

Key Findings

MRT fine-tuning improves Qwen2.5-7B-Instruct overall accuracy from 16.20% to 68.95%.
The results demonstrate the effectiveness of sequence-level optimization for domain-specific structured generation.
The approach addresses a practical utility-domain problem: converting heterogeneous outage reports into standards-compliant XML (CIM IEC 61968-3).

Significance

The work shows that relatively small language models can be substantially enhanced for specialized structured-generation tasks by optimizing task-level metrics directly, rather than relying on generic next-token prediction objectives.