Legal Domain LLM Training & Alignment
End-to-end training pipeline for legal domain LLMs. Achieved SOTA on contract analysis benchmarks using SFT, DPO, and task-specific LoRA adapters.
Overview
Built and operated the full training and alignment pipeline for legal domain LLMs at BHSN. Scope includes dataset curation, SFT, preference optimization, and multi-task evaluation across contract review, clause extraction, and document analysis.
Key Achievements
Legal Domain Adaptive Continuous Pretraining (2024 Q1)
- Built a reusable CPT recipe applicable across decoder LLMs (Llama, Solar, etc.) and curated a ~100GB legal-domain corpus from scratch
- Combined LoRA with pretraining to perform Legal Domain Adaptive CPT while preventing excessive distribution shift and overfitting
- Consistent gains over base on an in-house legal benchmark suite (contract-type classification, civil/criminal case-type classification, policy review, legal-term multiple choice, bar-exam items)
- General capability indicators such as MMLU remained stable while target-domain metrics rose selectively — domain specialization without general-performance regression
- Laid the foundation for the subsequent 2024 Q2 – 2025 multi-task SFT stage
Contract Analysis SOTA
- Multi-task SFT model achieved overall 0.9138 — +4.7% over GPT-4.1 (0.8731)
- Party identification: 97.5%, query answerability: 95.75%
- Checklist model v2 human eval 80.41% (on par with or above GPT-4o at 79.16%)
Long-Context Performance
- Fine-tuned models for processing 200K-character EPC contracts
- Hit rate: base Gemma-3 0.081 → SFT 0.383 (approaching GPT-4.1 at 0.390)
Task-Specific Adapters
- Trained 5 LoRA adapters on Qwen2.5-14B backbone: checklist review, redline revision, translation, label extraction, Contract-AI
- Adapter-based serving enabling simultaneous on-premise deployment and commercial model replacement
Pipeline & Infrastructure
- Dataset operations across 13+ repositories (systematic version control)
- Evaluation framework: traditional metrics + LLM-as-a-Judge + legal expert rubrics
- Training on AICA-provided H100x8 GPU infrastructure
- 10+ model benchmark comparisons (GPT-4.1, Claude Sonnet, Gemini, Qwen3, Gemma-3)
Academic Context
Task-specific LoRA adapter training extends the low-rank adaptation technique from LoRA (Hu et al., ICLR 2022) to legal domain multi-task settings. For preference optimization, DPO (Rafailov et al., NeurIPS 2023) was applied — directly optimizing policy with classification loss without a reward model — ensuring training stability and implementation simplicity. On the memory-efficient training front, quantization techniques from QLoRA (Dettmers et al., NeurIPS 2023) were also referenced for operating large-scale model training on H100/A100 infrastructure.
The project applies a layered CPT → SFT → DPO stack, combining LoRA at each stage to control catastrophic forgetting and overfitting. Legal Domain Adaptive CPT in 2024 Q1 established domain pre-exposure, and the subsequent multi-task SFT reached overall 0.9138 — surpassing the GPT-4.1 SOTA (0.8731) in a resource-efficient manner. The result demonstrates that hierarchical adaptation, not any single technique, is the effective path for legal AI.
Tech Stack
PyTorch, Transformers, PEFT/LoRA, vLLM, Weights & Biases, H100/A100 GPU