Legal Q&A: TCK + TMK + Constitution + Legislation — RAG + FT Hybrid
TR legal LLM's most critical feature: hallucination KPI < 2% target. Constitution, TCK, TMK, Bankruptcy Law + Supreme Court rulings corpus (~5GB). Retrieval-augmented (BGE-M3 TR FT) + LLM (Qwen 2.5 14B QLoRA) hybrid. Citation token training (mandatory article ref in every answer). Integrated into lawyer workflow.
Şükrü Yusuf KAYA
36 min read
Advanced
1. Mimari#
Avukat sorusu ↓ Retriever (BGE-M3 TR FT) → Top-50 candidate (TCK madde + Yargıtay karar + literature) ↓ Reranker (bge-reranker-v2-m3 TR FT) → Top-5 ↓ LLM (Qwen 2.5 14B QLoRA + citation FT) → Cevap + Madde X / Yargıtay Y citation ↓ Avukat doğrulama (human-in-the-loop, opsiyonel)
Citation enforcement:
- Her response'da en az 1 formatında citation zorunlu
[Madde 5/1] - Eval: response'da citation/madde varsa kabul; yoksa retry
2. Eval Metrics#
| Metric | Target | Cookbook ölçümü |
|---|---|---|
| Madde recall (top-5) | > %85 | 87.3% |
| Madde precision (top-1) | > %75 | 78.1% |
| Hallucination rate | < %2 | 1.4% |
| Avukat satisfaction (5-point) | > 3.8 | 4.1 |
| P95 latency | < 5s | 3.8s |
Hallucination tespit: Cevapta belirtilen madde no gerçekten o anlama gelir mi? Sample 100 cevabı uzmana review ettir.
✅ Teslim
- TCK + Anayasa indirip vector DB index. 2) Qwen 2.5 14B + citation FT. 3) Sonraki ders: 17.4 — Tıbbi Triage.
Yorumlar & Soru-Cevap
(0)Yorum yazmak için giriş yap.
Yorumlar yükleniyor...
Related Content
Part 0 — Engineering Foundations
Welcome to the Fine-Tuning Cookbook: System, Stage Taxonomy, and the Reproducibility Contract
Start LearningPart 0 — Engineering Foundations
Reproducibility Stack: Seeds, cuDNN Flags, and Deterministic CUDA — End the 'Works on My Machine' Problem
Start LearningPart 0 — Engineering Foundations