How We Cut Target Leakage from 92% to 0.1% in Court Outcome Prediction (and Built a Triple-A MCP Server)
说明与用途
我们通过清理文本将法院判决预测中的目标泄漏从92%降至0.1%,并使用LightGBM训练模型,获得78.2%的准确率。模型已作为零依赖的MCP服务器部署。
问题
AI models for court outcome prediction suffer from target leakage, where 92% of raw texts contain the outcome verbatim, causing models to simply read back the answer instead of learning legal reasoning.
结果
Target leakage dropped from 92% to 0.1%; model achieved 78.2% accuracy (vs 43.7% baseline) with macro‑F1 of 77.1% across 609,715 cases. MCP server received Triple‑A audit rating.
工作流
- 1. Implement strict sanitization: strip dictum, summary lines, and outcome‑announcing phrases from raw texts.
- 2. Continuously measure residual outcome markers to confirm leakage reduction.
- 3. Train LightGBM on the sanitized dataset.
- 4. Evaluate with 5‑fold cross‑validation on out‑of‑fold data.
- 5. Wrap the pipeline into a zero‑dependency MCP server exposing three tools: rechtspraak_cijfers, lekkage_check, voorspel_uitkomst.
证据
Leakage dropped from 92% to 0.1% (around 1 in 1,000 texts).
Overall Accuracy: 78.2% (against a 43.7% majority baseline)
Zero external dependencies: Single Python file using only standard library
- 记录日期: 2026-09-15T21:00:45.174Z