Transforming Business with AI

Fine-tuned LLM on 60K+ legal contracts, improving accuracy by 28% while reducing costs by 62%.

Challenge

The platform used GPT-4 via API for contract analysis but found general-purpose models insufficiently accurate on legal clause detection and compliance language. Cost and latency were also concerns.

Solution

• 1
  Fine-tuned a GPT-J 6B model on 60K+ annotated contract samples
• 2
  Applied instruction-tuning with RLHF techniques on lawyer-generated feedback
• 3
  Optimized tokenization to reduce latency and inference cost

Execution

• Used Hugging Face's Trainer with PEFT (LoRA) for cost-efficient fine-tuning
• Built an internal annotation platform with active learning loop
• Benchmarked results against GPT-4 and Claude with blind evaluations

Results

Key Metrics